{ "meta": { "owner": "Bhavani Adula", "generated": "2026-09-25", "last_full_verification": "2026-09-25", "verification_method": "Direct inspection of locally cloned repositories at pinned commit SHAs, cross-checked against resume PDFs and LinkedIn export. No claim is marked VERIFIED on the basis of a README, a resume line, or a generated description alone.", "states": { "VERIFIED": "Checked against primary source evidence (code, committed data file, or test) and the evidence supports the claim as worded.", "VERIFICATION_REQUIRED": "Plausible and self-reported but no public primary source exists. Typically employer-internal. Never stated as established fact.", "UNSUPPORTED": "Primary source evidence was inspected and does not support the claim as worded.", "DEPRECATED": "Previously made and since retracted or superseded." }, "repos": { "sfd": { "name": "silent-failure-detection", "url": "https://github.com/Bhavani-abhavi/silent-failure-detection", "sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "provenance": "original", "author": "Bhavani Adula" }, "cliniq": { "name": "ClinIQ-AI-for-Substance-Abuse-Risk-Detection", "url": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection", "sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "provenance": "team project", "author": "Bhavani Adula and team (built together)" }, "churn": { "name": "CUSTOMER-CHURN-PREDICTION-USING-LLMs-WITH-SHAP-AND-BEHAVIOR-SUMMARIES", "url": "https://github.com/Bhavani-abhavi/CUSTOMER-CHURN-PREDICTION-USING-LLMs-WITH-SHAP-AND-BEHAVIOR-SUMMARIES", "sha": "6e0847212820a243dc9dfd421df430e02bac02d9", "provenance": "team course project", "author": "team 'Data Sorcerers', UMKC" }, "sssd": { "name": "SS-SD", "url": "https://github.com/Bhavani-abhavi/SS-SD", "sha": "c5ba67c0cfa7f5a9cc8b5a031c4175d687c6483f", "provenance": "fork", "upstream": "https://github.com/ango3636/SS-SD", "author": "team capstone; upstream commits authored by teammate" }, "meal": { "name": "Smart-Meal-Planner-Using-LLM", "url": "https://github.com/Bhavani-abhavi/Smart-Meal-Planner-Using-LLM", "sha": "8bad9b3a671b9efa5c0558390382a6b9fc8cf945", "provenance": "original", "author": "Bhavani Adula" }, "nas": { "name": "BUILDING-WEB-SERVER-FOR-NAS-SERVERS", "url": "https://github.com/Bhavani-abhavi/BUILDING-WEB-SERVER-FOR-NAS-SERVERS", "sha": "b455f3bdf960bc1c56d4c38dd4a79af3a5720db5", "provenance": "original", "author": "Bhavani Adula" }, "fundproof": { "name": "Fundproof", "url": "https://github.com/Bhavani-abhavi/Fundproof", "sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "provenance": "fork of a two-person joint project", "author": "two-person team; specs, build prompts and evaluation protocol co-written by Bhavani Adula; implemented by Claude Code" } } }, "claims": [ { "id": "sfd.scale.code", "claim": "The silent-failure-detection framework is 8,209 lines of Python across 56 modules with 261 tests.", "verification_status": "VERIFIED", "category": "engineering_scale", "source_type": "code", "source_location": "repo-wide count at pinned SHA", "verification_note": "find . -name '*.py' | xargs wc -l gives 8209 total across 56 files; 261 functions match ^def test_ under tests/.", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/tree/1a517a8b5275874dd677f84617b419cdaeaed4ac", "artifact_link": null, "public_safe": true, "tags": [ "python", "testing", "ml-engineering", "scale" ], "last_verified": "2026-09-25" }, { "id": "sfd.design.label_isolation", "claim": "The performance estimator is structurally prevented from reading the labels it is scored against, enforced at the import graph by an import-linter contract and again by pytest.", "verification_status": "VERIFIED", "category": "system_design", "source_type": "code", "source_location": "pyproject.toml, contract 'estimation cannot reach labels'", "verification_note": "type=forbidden, source_modules=['estimation'], forbidden_modules=['domains','backtest','reports']. The inline comment gives the reasoning: reaching a domain adapter would expose default_label_within_horizon and invalidate the evaluation. A second contract keeps drift_core domain-agnostic. Mirrored by tests/test_module_boundaries.py (153 lines).", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/pyproject.toml", "artifact_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/tests/test_module_boundaries.py", "public_safe": true, "tags": [ "system-design", "testing", "evaluation-integrity", "architecture", "ml-engineering" ], "last_verified": "2026-09-25" }, { "id": "sfd.result.backtest_scale", "claim": "A credit-default model frozen after training on 2013 H1 was backtested across 35 monthly windows covering 1,054,948 loans (2014-01 to 2016-11).", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "code_and_artifact", "source_location": "README.md header; reports/backtest/signals.csv contains exactly 35 window rows", "verification_note": "signals.csv row count is 35, window_id runs from 2014-01. Loan count stated in README and consistent with the 24-month fixed-horizon cut described there, under which 1,285,664 rows survive.", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/scripts/run_backtest.py", "artifact_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/reports/backtest/signals.csv", "public_safe": true, "tags": [ "ml-engineering", "backtesting", "monitoring", "data-engineering" ], "last_verified": "2026-09-25" }, { "id": "sfd.result.silent_failure", "claim": "AUC held flat (0.6719 reference to 0.6661 final) while Brier rose 17% relative and the calibration gap widened 27-fold, so AUC-based monitoring would have reported nothing wrong.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "artifact", "source_location": "README.md section 3 table; reports/truth/true_performance_M.csv", "verification_note": "Reference 2013 H2: AUC 0.6719, Brier 0.0916, gap -0.0019. 2016-11: AUC 0.6661, Brier 0.1069, gap -0.0259. Worst window 2015-12: gap -0.0513, which is the 27x figure against the -0.0019 reference. At that point the model under-stated portfolio default risk by about 36% relative. The base rate rose 19% while mean predicted risk fell slightly, which rules out the explanation that applicants simply became riskier.", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/pipeline/calibration.py", "artifact_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/reports/truth/true_performance_M.csv", "public_safe": true, "tags": [ "ml-engineering", "calibration", "monitoring", "evaluation", "negative-result" ], "last_verified": "2026-09-25" }, { "id": "sfd.result.detectors_blind", "claim": "No unsupervised signal gave genuine early warning: of six detectors, three (KS, Wasserstein, multivariate) fired on 100% of windows and prediction drift never fired at all.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "artifact", "source_location": "README.md results table; reports/backtest/signals.csv and latency.csv", "verification_note": "Confirmed against signals.csv: ks_fired, wasserstein_fired and multivariate_fired are True on the first window and remain True; prediction_fired is False throughout. The README states that a detector firing on 100% of windows has maximum nominal lead time and zero information, and pre_onset_alert_rate exists in backtest/latency.py specifically to make that unquotable as a success.", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/backtest/latency.py", "artifact_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/reports/backtest/signals.csv", "public_safe": true, "tags": [ "evaluation", "negative-result", "monitoring", "research-methodology" ], "last_verified": "2026-09-25" }, { "id": "sfd.integrity.retraction", "claim": "The project publicly retracted its own headline finding after testing it: an originally reported anti-correlation (r = -0.38, p = 0.023) did not survive correction for serial dependence, and the corrected result is weaker.", "verification_status": "VERIFIED", "category": "research_integrity", "source_type": "code_and_artifact", "source_location": "README.md section 2; HEAD commit message 'Retract the anti-correlated claim; it does not survive verification'", "verification_note": "Three independent routes reported: effective n falls from 35 to 13.0, adjusted p = 0.195, moving-block bootstrap CI [-0.730, +0.280], and first differences flip the sign to r = +0.75. What survives is stated separately and is serial-correlation-robust: biased low in 35 of 35 windows (sign test p = 5.8e-11, still 2.4e-4 discounted to 13 effective observations), mean relative error -23.4%, Newey-West HAC slope on the gap +0.00096 per window (p = 0.0016). Reproducible via scripts/verify_correlation.py.", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/scripts/verify_correlation.py", "artifact_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/reports/backtest/correlation_significance.csv", "public_safe": true, "tags": [ "research-methodology", "statistics", "integrity", "evaluation" ], "last_verified": "2026-09-25" }, { "id": "sfd.failure.blinded_checks", "claim": "One bug pattern was found three times in code and twice more in the write-up: a safety check computed downstream of a transformation that removes the evidence the check looks for.", "verification_status": "VERIFIED", "category": "failure_case", "source_type": "code_and_artifact", "source_location": "README.md section 4; drift_core/validity.py; docs/findings_log.md", "verification_note": "Three code instances: (1) permutation floor, only 15 permutations so the smallest achievable p of 0.0625 sat above alpha and AUC 1.0 was reported as no drift; (2) all-NaN KS, NaN-dropping ran before any validity check leaving an empty comparison returning statistic=nan, is_drifted=False; (3) ESS guardrail, weight clipping at the 99th percentile capped exactly the large weights ESS exists to detect, returning a confident base rate of 0.534 on windows six standard deviations apart. A fourth instance was in the prose, not the code. A fifth occurred inside the fix for the fourth and produced p = 0, a value no finite test can produce, pinned by test_non_integer_override_never_yields_p_equal_zero. The response was structural rather than three separate fixes: one shared contract in drift_core/validity.py, every result carries a status distinct from its severity, and is_drifted=False is never sufficient on its own.", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/drift_core/validity.py", "artifact_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/docs/findings_log.md", "public_safe": true, "tags": [ "debugging", "testing", "failure-analysis", "system-design" ], "last_verified": "2026-09-25" }, { "id": "sfd.result.importance_weighting", "claim": "Importance weighting, which is valid under covariate shift, answered only 4 of 35 windows and was still 19% low on those; the reference-vs-current discriminator AUC ran 0.897 to 0.9998, leaving no common support to reweight across.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "code_and_artifact", "source_location": "README.md section 1; estimation/importance.py; reports/backtest/estimates.csv", "verification_note": "The framing is the measured outcome rather than a caveat: correcting for everything observable about P(X) left the error essentially intact, because the error was in P(Y|X).", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/estimation/importance.py", "artifact_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/reports/backtest/estimation_error_summary.csv", "public_safe": true, "tags": [ "ml-engineering", "statistics", "research-methodology", "negative-result" ], "last_verified": "2026-09-25" }, { "id": "sfd.result.fpr_zero", "claim": "Measured false-positive rate against a synthetic null was 0.0% across 560 feature-window tests per method, which makes the constant firing on real data a property of the data rather than miscalibrated detectors.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "artifact", "source_location": "README.md; reports/calibration/intrinsic_fpr.csv, null_fpr.csv, THRESHOLD_CALIBRATION.md", "verification_note": "Random splits of the healthy reference period. All four univariate detectors alerted zero times. The README reads this as making the result worse rather than better: the detectors are correctly calibrated and fire constantly because real credit populations never stop moving.", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/scripts/calibrate_lending_club.py", "artifact_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/reports/calibration/THRESHOLD_CALIBRATION.md", "public_safe": true, "tags": [ "evaluation", "statistics", "monitoring" ], "last_verified": "2026-09-25" }, { "id": "sfd.conclusion.architecture", "claim": "The design conclusion is that fixed-threshold alerting is the wrong architecture: the same statistics correlate with the true calibration gap (multivariate AUC r = 0.88, rising monotonically 0.636 to 0.803) but carry their signal in trajectory rather than level, so thresholding discards the informative part.", "verification_status": "VERIFIED", "category": "design_decision", "source_type": "artifact", "source_location": "README.md 'Design conclusion'", "verification_note": "MANDATORY QUALIFIER carried by the source itself: these correlations are computed on autocorrelated series, are descriptive, are offered without p-values, and the same serial-dependence correction that sank r = -0.38 would widen their intervals substantially. The monotonicity of the multivariate AUC is presented as the more robust observation. The caveat must travel with the number.", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/drift_core/multivariate.py", "artifact_link": null, "public_safe": true, "tags": [ "system-design", "monitoring", "design-decision" ], "last_verified": "2026-09-25" }, { "id": "sfd.limits.stated", "claim": "The project states its own limits: there was no healthy deployment period to detect across (33 of 35 windows breach the Brier threshold), prediction drift's 'never fired' is threshold-dependent, the onset definition is contestable and is swept rather than chosen, and results rest on a single dataset and model class.", "verification_status": "VERIFIED", "category": "research_integrity", "source_type": "code_and_artifact", "source_location": "README.md 'Honest limits'; reports/backtest/onset_sensitivity.csv; scripts/rescore_backtest.py", "verification_note": "Latency is reported under four different ground-truth metrics and both free parameters of the onset rule are swept rather than chosen. Max PSI on the score distribution was 0.0894 against a 0.1 gate; at a 0.05 gate it fires on 11 of 35 windows and still would not have led the onset. Degradation began about two months after the training window ended, so the maximum achievable lead time was 2-4 windows regardless of detector quality.", "project_or_experience": "silent-failure-detection", "commit_sha": "1a517a8b5275874dd677f84617b419cdaeaed4ac", "exact_code_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/scripts/rescore_backtest.py", "artifact_link": "https://github.com/Bhavani-abhavi/silent-failure-detection/blob/1a517a8b5275874dd677f84617b419cdaeaed4ac/reports/backtest/onset_sensitivity.csv", "public_safe": true, "tags": [ "research-methodology", "integrity", "limitations" ], "last_verified": "2026-09-25" }, { "id": "cliniq.scale.corpus", "claim": "ClinIQ analysed 52,184 patient-written drug reviews covering 2008-2017, of which 3,316 (6.4%) were SUD-relevant.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "artifact", "source_location": "outputs/pipeline_summary.txt; outputs/sud_clusters.csv has exactly 3,316 rows", "verification_note": "Row count confirmed programmatically. Source is the public Kaggle/UCI Drug Review dataset, not patient records. No PHI is involved.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/data/load_reviews.py", "artifact_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/outputs/pipeline_summary.txt", "public_safe": true, "tags": [ "nlp", "data-analysis", "healthcare" ], "last_verified": "2026-09-25" }, { "id": "cliniq.result.cohorts", "claim": "Clustered 3,316 SUD-flagged reviews into 37 behavioural cohorts using UMAP and HDBSCAN over 384-dimensional all-MiniLM-L6-v2 embeddings.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "artifact", "source_location": "outputs/sud_clusters.csv, column behavioral_cluster", "verification_note": "38 unique values including HDBSCAN's -1 noise label, so 37 real cohorts. Embedding dimension 384 corroborated by db/schema.sql declaring embedding vector(384). IMPORTANT CAVEAT added 2026-10-01: sud_clusters.csv still holds 3,316 rows and was NOT regenerated after the proxy-label correction, so roughly 24% of the clustered reviews are now known to be false positives from the 'meth' substring bug. The 37 cohorts are therefore a result over a contaminated set. Re-running the clustering on the 2,513 corrected reviews would settle whether the structure survives. See cliniq.integrity.label_correction.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/analysis/task2_temporal_behavioral.py", "artifact_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/outputs/sud_clusters.csv", "public_safe": true, "tags": [ "nlp", "embeddings", "clustering", "unsupervised" ], "last_verified": "2026-09-25" }, { "id": "cliniq.result.distress_trend", "claim": "Patient distress proportion rose from 1.7% in 2008 to 30.5% in 2017 across the review corpus.", "verification_status": "DEPRECATED", "category": "experimental_result", "source_type": "artifact", "source_location": "outputs/temporal_trends.csv, which the repository now states predates the label fix", "verification_note": "SUPERSEDED on 2026-09-29 and retained here so the record shows what changed. The SUD proxy label matched 'meth' as a SUBSTRING, so reviews of methylphenidate, sulfamethoxazole, methylprednisolone, dextromethorphan and indomethacin were counted as substance-use related. The repository states that 803 of 3,316 proxy-positive reviews, 24%, were positive only because of that. 'meth' is now required to be a whole word, pinned by test_meth_matches_only_as_a_whole_word. Recomputed on corrected labels the trend is 2.0% to 24.8%, not 1.7% to 30.5%, over 2,513 reviews rather than 3,316. The superseded figures must not be quoted. See cliniq.result.distress_trend_corrected.", "project_or_experience": "ClinIQ", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "nlp", "data-analysis", "healthcare" ], "last_verified": "2026-09-25" }, { "id": "cliniq.arch.three_tier", "claim": "ClinIQ compares three retrieval tiers on the same labelled task: rule-based, embedding cosine similarity, and Claude plus pgvector RAG.", "verification_status": "VERIFIED", "category": "architecture", "source_type": "code_and_artifact", "source_location": "analysis/task1_signal_detection.py; outputs/method_comparison_results.csv", "verification_note": "A genuine baseline ladder over a shared evaluation set, not three unrelated demos. pgvector confirmed at db/schema.sql line 1 (CREATE EXTENSION vector) and line 47 (embedding vector(384)). Anthropic SDK confirmed at agent/cliniq_agent.py line 8; the API key is read from the environment, not hardcoded.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/analysis/task1_signal_detection.py", "artifact_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/db/schema.sql", "public_safe": true, "tags": [ "baselines", "llm", "pgvector", "postgresql", "rag", "retrieval", "sql", "vector-search" ], "last_verified": "2026-09-25" }, { "id": "cliniq.result.method_comparison", "claim": "The three tiers trade off sharply against each other: rule-based reached precision 0.861 and recall 0.847, embedding reached precision 0.504 and recall 1.000, and LLM plus RAG reached precision 0.938 and recall 0.400.", "verification_status": "VERIFIED", "category": "evaluation", "source_type": "artifact", "source_location": "outputs/method_comparison_results.csv", "verification_note": "Exact committed values: rule_based 0.8610169491525423 / 0.8466666666666667 (tp 254, fp 41, fn 46, tn 259); embedding 0.5042016806722689 / 1.0 (tp 300, fp 295, fn 0, tn 5); llm_rag 0.9375 / 0.4 (tp 120, fp 8, fn 180, tn 292). Elapsed time spans three orders of magnitude: 0.05s, 9.3s, 1772.7s. CRITICAL: no single tier achieved both high precision and perfect recall. See CONFLICTS.md entry C-001. CAVEAT added 2026-10-01: the CSV is byte-identical at the new commit, so these numbers are unchanged, but the repository now states they were scored against the OLD labels and were not rerun after the correction, because doing so needs the database and a paid API. Re-run offline on corrected labels the rules baseline scores F1 0.846 rather than 0.854. Treat these as the original team benchmark rather than the current best estimate.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/analysis/task1_signal_detection.py", "artifact_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/outputs/method_comparison_results.csv", "public_safe": true, "tags": [ "evaluation", "rag", "retrieval", "baselines", "precision-recall" ], "last_verified": "2026-09-25" }, { "id": "cliniq.result.revenue_synthetic", "claim": "The pipeline identified 219 SUD documentation gaps carrying $537,340.22 of revenue at risk, computed from CMS FY2024 IPPS Final Rule DRG weights.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "artifact", "source_location": "outputs/pipeline_summary.txt, lines 36-41", "verification_note": "MANDATORY QUALIFIER: the same summary block states 'Synthetic claims analyzed: 220 (all fictional)'. The dollar figure is a modelled result over fictional claims demonstrating that the DRG-weight calculation works, not revenue identified in a real payer population. Any presentation omitting the word synthetic misrepresents it. See CONFLICTS.md entry C-002.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/run_pipeline.py", "artifact_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/outputs/pipeline_summary.txt", "public_safe": true, "tags": [ "healthcare", "domain-modelling", "synthetic-data" ], "last_verified": "2026-09-25" }, { "id": "cliniq.design.grounding", "claim": "Every ClinIQ output is grounded in a cited government source, with source_url stored per embedding so each answer is auditable back to its document.", "verification_status": "VERIFIED", "category": "design_decision", "source_type": "code", "source_location": "db/schema.sql lines 38-47; agent/cliniq_agent.py; agent/explainability.py", "verification_note": "The schema comment reads 'source_url field makes every embedding fully auditable'. The knowledge base is 122 embeddings over 10 registered source documents: 30 DRG-weight chunks from CMS FY2024 IPPS Final Rule Table 5, 32 ICD-10 chunks from the FY2024 official code set, and 60 SUD-signal chunks. pipeline_summary.txt confirms 'Every output cites RAG source: YES'.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/db/schema.sql", "artifact_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/agent/explainability.py", "public_safe": true, "tags": [ "auditability", "citations", "grounding", "healthcare", "postgresql", "rag", "safety", "sql" ], "last_verified": "2026-09-25" }, { "id": "cliniq.provenance", "claim": "ClinIQ was built for the NSF NRT Research-A-Thon 2026 (4th place, self-reported) and later extended with FHIR interoperability, a prior-authorization agent, an active-learning labelling workbench and a counterfactual bias test. The repository is public on this account.", "verification_status": "VERIFIED", "category": "provenance", "source_type": "repository_metadata", "source_location": "Repository README and commit history", "verification_note": "Built together as a team project. The README at the pinned commit documents each extension with its evaluation results.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection", "artifact_link": null, "public_safe": true, "tags": [ "provenance", "collaboration", "teamwork" ], "last_verified": "2026-09-25" }, { "id": "cliniq.award.nsf", "claim": "ClinIQ placed 4th at the NSF NRT Research-A-Thon 2026 at UMKC.", "verification_status": "VERIFICATION_REQUIRED", "category": "award", "source_type": "self_reported", "source_location": "Repository description; LinkedIn post with event poster", "verification_note": "No organiser-published results page was located during verification. The claim is consistent across the repo description, LinkedIn and both resumes, but consistency across self-authored sources is not independent confirmation. Presented as self-reported until an organiser result is linked.", "project_or_experience": "ClinIQ", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "award", "recognition" ], "last_verified": "2026-09-25" }, { "id": "churn.result.bert", "claim": "A fine-tuned BERT churn classifier reached 0.793 accuracy and macro F1 0.73 on 1,409 held-out customers.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "artifact", "source_location": "C_PDS_DATA_SORCERERS.ipynb, stored cell outputs", "verification_note": "The exact value in the committed notebook output is 0.7927608232789212, with the sklearn classification report showing macro avg 0.73 over support 1409. A second, weaker run at 0.77 accuracy and macro avg 0.68 is also present in the same notebook; the reported figure is the stronger of the two and that should be said if asked.", "project_or_experience": "Explainable Customer Churn Prediction", "commit_sha": "6e0847212820a243dc9dfd421df430e02bac02d9", "exact_code_link": "https://github.com/Bhavani-abhavi/CUSTOMER-CHURN-PREDICTION-USING-LLMs-WITH-SHAP-AND-BEHAVIOR-SUMMARIES/blob/6e0847212820a243dc9dfd421df430e02bac02d9/pds_data_sorcerers.py", "artifact_link": "https://github.com/Bhavani-abhavi/CUSTOMER-CHURN-PREDICTION-USING-LLMs-WITH-SHAP-AND-BEHAVIOR-SUMMARIES/blob/6e0847212820a243dc9dfd421df430e02bac02d9/C_PDS_DATA_SORCERERS.ipynb", "public_safe": true, "tags": [ "nlp", "bert", "fine-tuning", "classification", "pytorch" ], "last_verified": "2026-09-25" }, { "id": "churn.arch.t5_shap", "claim": "T5 was chained after the classifier to convert structured customer rows into plain-language risk narratives, and SHAP over XGBoost ranked the churn drivers as contract type, tenure and monthly charges.", "verification_status": "VERIFIED", "category": "architecture", "source_type": "code_and_artifact", "source_location": "C_PDS_DATA_SORCERERS.ipynb; customer_summaries_with_sentiment.csv", "verification_note": "1.7 MB of generated summaries are committed as an artifact rather than only described, so the model output is directly inspectable.", "project_or_experience": "Explainable Customer Churn Prediction", "commit_sha": "6e0847212820a243dc9dfd421df430e02bac02d9", "exact_code_link": "https://github.com/Bhavani-abhavi/CUSTOMER-CHURN-PREDICTION-USING-LLMs-WITH-SHAP-AND-BEHAVIOR-SUMMARIES/blob/6e0847212820a243dc9dfd421df430e02bac02d9/pds_data_sorcerers.py", "artifact_link": "https://github.com/Bhavani-abhavi/CUSTOMER-CHURN-PREDICTION-USING-LLMs-WITH-SHAP-AND-BEHAVIOR-SUMMARIES/blob/6e0847212820a243dc9dfd421df430e02bac02d9/customer_summaries_with_sentiment.csv", "public_safe": true, "tags": [ "nlp", "t5", "shap", "explainability", "xgboost" ], "last_verified": "2026-09-25" }, { "id": "sssd.arch.kinematic_encoder", "claim": "Stable Diffusion 1.5's CLIP text encoder was replaced with a custom KinematicEncoder mapping 76-dimensional robot kinematics into cross-attention, fine-tuned with LoRA rank 4 via peft.", "verification_status": "VERIFIED", "category": "architecture", "source_type": "code", "source_location": "src/suturing_pipeline/ at pinned SHA", "verification_note": "Verified in the fork's code: 76-dim input, MLP 256/512, producing 77x768 tokens, plus a gesture nn.Embedding(16,768) and LayerNorm. LoRA rank 4 applied to to_q, to_k, to_v and to_out.0 via peft. DDPM training, DDIM sampling, 256px, batch 4, 50 epochs, frame_stride 90, checkpoint step_1480. JIGSAWS Suturing loader uses OneTrialOut folds with StandardScaler and lazy OpenCV seeking.", "project_or_experience": "SS-SD", "commit_sha": "c5ba67c0cfa7f5a9cc8b5a031c4175d687c6483f", "exact_code_link": "https://github.com/Bhavani-abhavi/SS-SD/tree/c5ba67c0cfa7f5a9cc8b5a031c4175d687c6483f/src/suturing_pipeline", "artifact_link": null, "public_safe": true, "tags": [ "computer-vision", "diffusion", "lora", "pytorch", "generative", "fine-tuning" ], "last_verified": "2026-09-25" }, { "id": "sssd.eval.cross_pair_baseline", "claim": "Generation quality was evaluated against a cross-pair baseline in which each generated frame is also scored against a different real frame, separating genuine conditioning from scene-average reconstruction.", "verification_status": "VERIFIED", "category": "evaluation", "source_type": "code", "source_location": "scripts/metrics_on_grid.py; scripts/video_quality_metrics.py", "verification_note": "metrics_on_grid.py computes PSNR, SSIM, histogram chi-squared and edge IoU with a diagonal-versus-off-diagonal comparison. video_quality_metrics.py computes per-frame SSIM, Farneback optical-flow correlation and MAD flicker flags. The baseline design is the methodologically interesting part and it is real in the code. LIMITATION: it was run over only 4 pairs from a smoke test, which is too few to support a strong quantitative claim.", "project_or_experience": "SS-SD", "commit_sha": "c5ba67c0cfa7f5a9cc8b5a031c4175d687c6483f", "exact_code_link": "https://github.com/Bhavani-abhavi/SS-SD/blob/c5ba67c0cfa7f5a9cc8b5a031c4175d687c6483f/scripts/metrics_on_grid.py", "artifact_link": null, "public_safe": true, "tags": [ "computer-vision", "evaluation", "baselines", "methodology" ], "last_verified": "2026-09-25" }, { "id": "sssd.metrics.psnr_ssim", "claim": "Generated frames scored PSNR 19-22 dB and SSIM 0.61-0.74 against a VAE reconstruction ceiling of 24-27 dB and 0.82-0.91.", "verification_status": "VERIFICATION_REQUIRED", "category": "experimental_result", "source_type": "team_report", "source_location": "SS-SD_Report_4pg.pdf (team report, May 2026), held locally and not published in the repository", "verification_note": "No metric values are committed in either the fork or upstream, so these numbers cannot be reproduced from the public repository, and this claim therefore carries NO code link: linking metrics_on_grid.py here would imply the script evidences these figures, which it does not. The script's existence is a separate, VERIFIED claim (see sssd.eval.cross_pair_baseline). The stated basis for the numbers is a 12-sample held-out expert grid at checkpoint step_1480 in a team report held privately. Committing reports/metrics.json would move this to VERIFIED in one commit.", "project_or_experience": "SS-SD", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "computer-vision", "evaluation", "metrics" ], "last_verified": "2026-09-25" }, { "id": "sssd.metrics.fid", "claim": "FID was used to benchmark generated surgical video.", "verification_status": "VERIFICATION_REQUIRED", "category": "evaluation", "source_type": "code", "source_location": "scripts/compute_fid.py, present in the fork since the upstream sync (commit b5b2cb8)", "verification_note": "CORRECTED 2026-09-28. An earlier review recorded this as UNSUPPORTED on the basis that compute_fid.py existed only upstream. That was wrong: the script is present in the fork, and has been since the 21 Sept 2026 upstream sync. What is still absent is any FID VALUE. A repository-wide search for committed metric values returns nothing, in any JSON or CSV, for FID, PSNR or SSIM. So the tooling is real and the result is not published. The claim is therefore self-reported rather than unsupported, and carries no code link, because linking the script would imply it evidences a number it has never been shown to produce. Running it and committing the output would move this to VERIFIED. See CONFLICTS.md C-003.", "project_or_experience": "SS-SD", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "computer-vision", "evaluation", "metrics" ], "last_verified": "2026-09-25" }, { "id": "sssd.provenance", "claim": "SS-SD is a team capstone. The repository on this account is a fork; upstream commits are the teammate's.", "verification_status": "VERIFIED", "category": "provenance", "source_type": "repository_metadata", "source_location": "GitHub fork metadata, upstream ango3636/SS-SD", "verification_note": "The fork was synced with upstream on 2026-09-21 and is current rather than a stale snapshot. 33 tests in the fork: 30 pass, 3 fail on Windows only, due to a NamedTemporaryFile unlink issue in test_jigsaws_metafile.py. Five scripts hard-code a teammate's local paths, notebooks 01-04 are empty stubs, and ControlNet is a todo scaffold. None of that is hidden.", "project_or_experience": "SS-SD", "commit_sha": "c5ba67c0cfa7f5a9cc8b5a031c4175d687c6483f", "exact_code_link": "https://github.com/Bhavani-abhavi/SS-SD", "artifact_link": null, "public_safe": true, "tags": [ "provenance", "collaboration" ], "last_verified": "2026-09-25" }, { "id": "hippocratic.eval.scenarios", "claim": "Developed and improved production healthcare voice AI agents built on Hippocratic AI's Polaris constellation architecture, supporting post-discharge patient conversations across a multi-hospital health system; ran root-cause analysis on 300+ flagged production conversations and built 240+ LLM evaluation scenarios and regression datasets from those failure patterns.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience", "source_type": "self_reported", "source_location": "Resume (Oct 2026); LinkedIn experience entry; role description supplied by Bhavani on 2026-10-06", "verification_note": "Employer-internal work at Hippocratic AI. No public artifact exists and none should, since the underlying material is production healthcare conversation data. This is normal for industry work and is not a mark against the claim, but it must be presented as self-reported and must never carry a code link. The adjacent public evidence for the same skill set is silent-failure-detection (evaluation design, regression testing, 261 tests) and ClinIQ (grounding and citation enforcement).", "project_or_experience": "Hippocratic AI", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "llm-evaluation", "eval-datasets", "regression-testing", "self-reported" ], "last_verified": "2026-10-06" }, { "id": "hippocratic.eval.regressions", "claim": "The regression suite caught 8 regressions pre-release and cut repeat failures in the largest category by 40%.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience_metric", "source_type": "self_reported", "source_location": "Resume (both variants); LinkedIn", "verification_note": "Employer-internal metric, not independently verifiable. Present as self-reported; never as an established number.", "project_or_experience": "Hippocratic AI", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "llm-evaluation", "regression-testing", "self-reported" ], "last_verified": "2026-09-25" }, { "id": "hippocratic.reliability.hardening", "claim": "Improved RAG and retrieval (document chunking, embeddings, semantic retrieval, reranking, grounding checks), tool calling (structured argument validation, failure handling), clinical escalation and fallback logic, and the real-time ASR-LLM-TTS voice pipeline: document-grounded response accuracy up 12%, invalid tool calls down 70%, missed escalations down 45%, p95 latency down 25% with no measurable regression in quality or safety.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience_metric", "source_type": "self_reported", "source_location": "Resume (Oct 2026); LinkedIn; role description supplied by Bhavani on 2026-10-06", "verification_note": "Four employer-internal percentage metrics, not independently verifiable. The underlying capability areas (retrieval grounding, tool-call argument validation, escalation and fallback design) are real and worth interviewing on, but each number must be attributed as self-reported every time it appears.", "project_or_experience": "Hippocratic AI", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "agent-reliability", "tool-calling", "grounding", "latency", "self-reported" ], "last_verified": "2026-10-06" }, { "id": "yrall.pipelines", "claim": "Built and maintained 12+ production ETL/ELT pipelines in Python and SQL processing 1.5M+ records daily from 8+ sources, deployed on Google Cloud Platform with BigQuery and Cloud Storage and orchestrated as Airflow DAGs with scheduling, retries, incremental loading and monitoring.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience", "source_type": "self_reported", "source_location": "Resume (both variants); LinkedIn experience entry", "verification_note": "Employer-internal work at Yrall Media Solutions, Jan 2023 to Jun 2024. No public repository. FIGURES REVISED 2026-10-02 by the owner, who states the earlier description understated the role: daily volume goes from 20K-50K to 1.5M+, and GCP, BigQuery and Cloud Storage are named for the first time. Both versions are self-reported and neither is publicly checkable, so the revision is recorded here rather than silently applied. LinkedIn still carries the earlier figure and needs the same edit, otherwise the two sources disagree.", "project_or_experience": "Yrall Media Solutions", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "data-engineering", "airflow", "etl", "postgresql", "sql", "gcp", "bigquery", "cloud", "self-reported" ], "last_verified": "2026-09-25" }, { "id": "yrall.lineage_ci", "claim": "Optimised SQL and transformation logic for roughly 28% faster pipeline execution, added record-level lineage and CI data-quality gates reducing recurring data issues by roughly 25%, and modelled dimensional reporting datasets over 5M+ historical records powering Power BI for 200+ campaigns.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience", "source_type": "self_reported", "source_location": "Resume (both variants)", "verification_note": "Employer-internal, self-reported, revised 2026-10-02 alongside yrall.pipelines. The percentage improvements are approximate on the resume and are presented that way here. Notable because the same instinct, make the output traceable and fail the build when it drifts, is independently demonstrated in public code: ClinIQ stores source_url per embedding and silent-failure-detection enforces module boundaries in CI. That is adjacent corroboration of the practice, not verification of the employer claim.", "project_or_experience": "Yrall Media Solutions", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "data-engineering", "dbt", "lineage", "ci", "sql", "data-quality", "self-reported" ], "last_verified": "2026-09-25" }, { "id": "skill.docker", "claim": "Docker and docker-compose are used to containerise a working project.", "verification_status": "VERIFIED", "category": "skill", "source_type": "code", "source_location": "ClinIQ Dockerfile and docker-compose.yml", "verification_note": "Present and part of the documented run path alongside a Makefile.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/Dockerfile", "artifact_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/docker-compose.yml", "public_safe": true, "tags": [ "docker", "devops", "containerisation" ], "last_verified": "2026-09-25" }, { "id": "skill.ci", "claim": "CI is configured with GitHub Actions, running tests on push and pull request.", "verification_status": "VERIFIED", "category": "skill", "source_type": "code", "source_location": "ClinIQ .github/workflows/ci.yml", "verification_note": "Triggers on push to main and develop and on pull_request to main; Python 3.11 matrix; checkout and setup-python steps confirmed.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/.github/workflows/ci.yml", "artifact_link": null, "public_safe": true, "tags": [ "ci", "github-actions", "testing", "devops" ], "last_verified": "2026-09-25" }, { "id": "skill.gemini", "claim": "The Smart Meal Planner calls the Google Gemini API, not Hugging Face Transformers.", "verification_status": "VERIFIED", "category": "skill", "source_type": "code", "source_location": "code.py lines 5-9", "verification_note": "API_URL targets generativelanguage.googleapis.com/v1beta2/models/gemini-1.5:generateText. The key is read from an environment variable and is not hardcoded, which is correct practice. Both the LinkedIn project description and the resume summary attribute this project to Hugging Face Transformers, which the code contradicts. See CONFLICTS.md entry C-004.", "project_or_experience": "Smart Meal Planner", "commit_sha": "8bad9b3a671b9efa5c0558390382a6b9fc8cf945", "exact_code_link": "https://github.com/Bhavani-abhavi/Smart-Meal-Planner-Using-LLM/blob/8bad9b3a671b9efa5c0558390382a6b9fc8cf945/code.py", "artifact_link": null, "public_safe": true, "tags": [ "llm", "api-integration", "gemini" ], "last_verified": "2026-09-25" }, { "id": "skill.nas_docs_only", "claim": "The NAS web server project (Apache/PHP/MariaDB portal with role-based access control) is published as documentation only.", "verification_status": "VERIFIED", "category": "skill", "source_type": "repository_metadata", "source_location": "Repository contains NAS_SERVERS.pptx and README.md only", "verification_note": "No implementation code is published, so any claim about the implementation is currently unverifiable from the public record. The repository description describes uploads, downloads, RBAC, authentication and an admin dashboard; none of that can be inspected.", "project_or_experience": "NAS Web Server", "commit_sha": "b455f3bdf960bc1c56d4c38dd4a79af3a5720db5", "exact_code_link": "https://github.com/Bhavani-abhavi/BUILDING-WEB-SERVER-FOR-NAS-SERVERS", "artifact_link": null, "public_safe": true, "tags": [ "web", "gap", "documentation-only" ], "last_verified": "2026-09-25" }, { "id": "gap.kubernetes", "claim": "Production Kubernetes ownership.", "verification_status": "UNSUPPORTED", "category": "gap", "source_type": "absence_of_evidence", "source_location": "Full repository sweep at pinned SHAs; no manifests, Helm charts or references found", "verification_note": "Not claimed on the resume or LinkedIn, and not demonstrated. What IS demonstrated adjacently: Docker and docker-compose (ClinIQ), GitHub Actions CI (ClinIQ), and a packaged Python project with enforced module boundaries (silent-failure-detection).", "project_or_experience": null, "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "gap", "kubernetes", "infrastructure" ], "last_verified": "2026-09-25" }, { "id": "gap.aws_public", "claim": "Hands-on AWS cloud engineering.", "verification_status": "VERIFICATION_REQUIRED", "category": "gap", "source_type": "self_reported", "source_location": "LinkedIn top skills lists AWS; the 2024 stock-prediction project describes cloud-based data management but has no public repository", "verification_note": "No AWS code, infrastructure definition or deployment artifact appears in any public repository. The AIML resume variant does not list AWS under MLOps and Cloud (it lists Docker, GitHub Actions, CI/CD, pytest, Streamlit, Git), which is the more defensible position. Either publish the stock-prediction work or treat AWS as coursework-level.", "project_or_experience": null, "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "gap", "aws", "cloud" ], "last_verified": "2026-09-25" }, { "id": "gap.serving_latency", "claim": "Production model serving, load testing, or horizontal scaling under real traffic.", "verification_status": "UNSUPPORTED", "category": "gap", "source_type": "absence_of_evidence", "source_location": "Repository sweep", "verification_note": "The p95 latency reduction at Hippocratic AI is self-reported and concerns agent latency rather than infrastructure scaling. No public load test, benchmark harness or serving deployment exists. This is a genuine gap, and the STRESS THIS ARCHITECTURE views are explicitly design proposals rather than evidence.", "project_or_experience": null, "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "gap", "scaling", "serving", "performance" ], "last_verified": "2026-09-25" }, { "id": "gap.agent_frameworks", "claim": "Production use of agent frameworks such as LangChain, LlamaIndex or LangGraph.", "verification_status": "UNSUPPORTED", "category": "gap", "source_type": "absence_of_evidence", "source_location": "Repository sweep at pinned SHAs, re-checked 2026-10-26", "verification_note": "UPDATED 2026-10-26. This note used to say that no agent framework appears in any public repository, which is no longer accurate: ClinIQ's later prior-authorization extension imports LangGraph (prior_auth/workflow.py at the pinned commit), and Fundproof, a two-person project implemented by Claude Code from specs Bhavani co-wrote, contains a LangGraph agent (fundproof.agent.checkpoint). Both run on synthetic data only and neither is deployed, so production use of an agent framework is still not demonstrated. The original ClinIQ retrieval pipeline calls the Anthropic SDK directly. Agentic-workflow claims from employer work rest on the self-reported Hippocratic AI entries.", "project_or_experience": null, "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "gap", "agents", "frameworks", "langgraph", "langchain", "llamaindex" ], "last_verified": "2026-10-26" }, { "id": "edu.masters", "claim": "M.S. in Computer Science, University of Missouri-Kansas City, completed May 2026.", "verification_status": "VERIFICATION_REQUIRED", "category": "education", "source_type": "self_reported", "source_location": "Resume (both variants); LinkedIn education entry", "verification_note": "Institution, programme name and completion date now agree across the resume and LinkedIn (LinkedIn updated 2026-09-25 from 'Data Science Analytics' to 'Computer Science', resolving CONFLICTS.md C-005). Like any credential this remains transcript-verifiable rather than publicly verifiable, so it is reported as self-reported, which is the normal state for a degree. Agreement between an owner's own sources is consistency, not independent confirmation; the registrar's record governs.", "project_or_experience": null, "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "education" ], "last_verified": "2026-09-25" }, { "id": "edu.bachelors", "claim": "B.Tech in Information Technology from J.B. Institute of Technology, May 2023.", "verification_status": "VERIFICATION_REQUIRED", "category": "education", "source_type": "self_reported", "source_location": "Resume (both variants)", "verification_note": "Not visible in the LinkedIn export reviewed, which shows UMKC and upGrad and indicates a third education entry that was not expanded. A standard credential, verifiable by transcript but not independently checkable here.", "project_or_experience": null, "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "education" ], "last_verified": "2026-09-25" }, { "id": "cert.list", "claim": "Certifications: Google Data Analytics, Data Science with AI (upGrad), Tableau for Data Science, Python (HackerRank).", "verification_status": "VERIFICATION_REQUIRED", "category": "certification", "source_type": "self_reported", "source_location": "Resume; LinkedIn licences section lists 6 certifications with credential links", "verification_note": "LinkedIn shows 'Show credential' links for at least the two Grow with Google entries: Prepare Data for Exploration (May 2024) and Ask Questions to Make Data-Driven Decisions (Dec 2023). Those are individual courses within the Google Data Analytics certificate rather than the full certificate; whether the full programme was completed should be confirmed before claiming it as one credential. See CONFLICTS.md entry C-006.", "project_or_experience": null, "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "certification" ], "last_verified": "2026-09-25" }, { "id": "leadership.ambassador", "claim": "Graduate Student Ambassador, UMKC School of Science and Engineering, Aug 2025 to May 2026; segmented a roughly 300-student international cohort in Excel and ran a student interview series.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience", "source_type": "self_reported", "source_location": "Resume; LinkedIn experience entry; LinkedIn post announcing the selection", "verification_note": "The role is corroborated by the LinkedIn experience entry and by a self-authored announcement post. The end date was corrected to May 2026 by the owner on 2026-09-29; LinkedIn still shows Jul 2026 and needs the same edit. The cohort-analysis detail is self-reported.", "project_or_experience": null, "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "leadership", "communication", "self-reported" ], "last_verified": "2026-09-25" }, { "verification_status": "VERIFICATION_REQUIRED", "source_type": "private_working_tree", "project_or_experience": "HoldLine", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "last_verified": "2026-10-01", "id": "holdline.status", "claim": "HoldLine is an in-progress build: a grounded voice agent that places real, time-limited holds on bakery inventory over the phone.", "category": "project_status", "source_location": "Private working tree, 30 commits, tags v0.1-foundation through v0.4-interfaces, HEAD 567b92ed67ac", "verification_note": "ACTIVE BUILD, NOT FINISHED. Four of nine planned phases are complete as of 2026-09-30. The phone layer and the LangGraph port are not built yet. The repository is private and has no GitHub remote, so nothing here is publicly inspectable and every figure below is VERIFICATION_REQUIRED by the rule this portfolio applies to everything else. The figures were checked directly against the working tree at the commit above rather than taken from a resume line, which is why they are stated at all, but a reader cannot check them and should not be asked to take them on trust. 13,138 lines of Python across 22 test files.", "tags": [ "in-progress", "backend", "python", "fastapi", "llm", "agents", "guardrails", "testing", "sqlite" ] }, { "verification_status": "VERIFICATION_REQUIRED", "source_type": "private_working_tree", "project_or_experience": "HoldLine", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "last_verified": "2026-10-01", "id": "holdline.design.post_generation_validation", "claim": "Every assistant response is validated after generation against the tool results from that turn, and violations are tiered by consequence: quantity claims retry once, allergen claims hard-block to a human.", "category": "design_decision", "source_location": "src/holdline/guardrails.py; tests/test_guardrails.py", "verification_note": "Verified by reading the source at the stated commit. The two tiers are explicit in the module: tier 1 carries one retry, tier 2 is never retried and never softened. The reasoning given in the file is that the consequences are not comparable, which is the design point worth asking about. Not publicly inspectable.", "tags": [ "in-progress", "guardrails", "llm", "grounding", "safety", "agent-reliability", "design-decision" ] }, { "verification_status": "VERIFICATION_REQUIRED", "source_type": "private_working_tree", "project_or_experience": "HoldLine", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "last_verified": "2026-10-01", "id": "holdline.testing.suite", "claim": "520 tests, of which 96 are guardrail tests covering false-positive categories, plus a concurrency test in which exactly one of ten simultaneous requests acquires the last unit, repeated 20 times.", "category": "engineering_scale", "source_location": "pytest collection; tests/test_guardrails.py; tests/test_concurrency.py", "verification_note": "Counted by running pytest --collect-only at the stated commit: 520 collected. test_guardrails.py holds 60 test functions that parametrise to 96 cases. test_concurrency.py sets REPEATS = 20 and races ten threads, asserting one winner, nine refusals carrying the reason 'insufficient', and exactly one new row in the ledger. The third assertion is the real one, since the first two could both hold while the database contained two holds. Not publicly inspectable.", "tags": [ "in-progress", "testing", "concurrency", "transactions", "guardrails" ] }, { "verification_status": "VERIFICATION_REQUIRED", "source_type": "private_working_tree", "project_or_experience": "HoldLine", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "last_verified": "2026-10-01", "id": "holdline.failure.live_model_pass", "claim": "Testing against the live model surfaced four defects that 444 passing offline tests had not, three of which were structurally invisible offline.", "category": "failure_case", "source_location": "DEFECTS.md entries 7 to 10", "verification_note": "Verified by reading DEFECTS.md at the stated commit; entries 7 to 10 are present and each records symptom, cause and fix. Examples: one schema keyword broke every conversation, the phrase 'and no allergies' escalated as an allergy, and 'twenty minutes' was parsed as containing a nut allergy. The file's closing section argues for a live pass at the end of every phase. Not publicly inspectable.", "tags": [ "in-progress", "failure-analysis", "llm-evaluation", "testing", "debugging" ] }, { "id": "cliniq.result.distress_trend_corrected", "claim": "On corrected labels, the share of SUD reviews rated 3 or below out of 10 rose from 2.0% in 2008 to 24.8% in 2017, across 2,513 reviews, while the opioid share of those reviews fell from 62% to 33%.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "artifact", "source_location": "outputs/temporal_corrected.json; analysis/temporal_check.py", "verification_note": "Read directly from the committed JSON at the pinned commit: 2008 distress_share 0.0201 and opioid_share 0.6242; 2017 distress_share 0.2481 and opioid_share 0.3342; sud_reviews 2,513. Distress is defined in the file as a rating of 3 or below out of 10, which is a rating threshold rather than a clinical measure and should be described that way. This replaces the earlier 1.7% to 30.5% figure, which rested on a substring-matching bug. The correction was made by the teammate who authored the repository's code commits.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/analysis/temporal_check.py", "artifact_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/outputs/temporal_corrected.json", "public_safe": true, "tags": [ "nlp", "data-analysis", "healthcare", "correction" ], "last_verified": "2026-10-01" }, { "id": "cliniq.integrity.label_correction", "claim": "A human labeling pilot exposed a substring-matching bug in the SUD proxy label: 803 of 3,316 proxy-positive reviews, 24%, qualified only because 'meth' matched inside unrelated drug names. Every offline result was rerun on corrected labels.", "verification_status": "VERIFIED", "category": "research_integrity", "source_type": "code_and_artifact", "source_location": "README.md 'Proxy-label correction (September 29, 2026)'; analysis/sud_labels.py; tests/test_bert_splits.py", "verification_note": "The bug was found because the labeling pilot disagreed with the proxy far more than expected and someone read the disagreements. The fix is pinned by test_meth_matches_only_as_a_whole_word; methamphetamine and methadone still match. The README states which results were rerun and which were not: the original team benchmark, rules F1 0.854 and LLM+RAG precision 0.938, needs the database and a paid API and so was left scored against the old labels. Authored by the teammate, not by Bhavani, and listed here because it changes the standing of figures she reported.", "project_or_experience": "ClinIQ", "commit_sha": "a701a223d66aba9d938f1f616f1237b8e8a1d76a", "exact_code_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/analysis/sud_labels.py", "artifact_link": "https://github.com/Bhavani-abhavi/ClinIQ-AI-for-Substance-Abuse-Risk-Detection/blob/a701a223d66aba9d938f1f616f1237b8e8a1d76a/tests/test_bert_splits.py", "public_safe": true, "tags": [ "research-methodology", "integrity", "evaluation", "labeling", "nlp" ], "last_verified": "2026-10-01" }, { "verification_status": "VERIFICATION_REQUIRED", "source_type": "self_reported", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "last_verified": "2026-10-02", "id": "movate.analyst", "claim": "Data Analyst Intern at Movate, Nov 2021 to Dec 2022: extracted and integrated CRM and internal system data with SQL into unified reporting datasets with quality checks, and delivered Tableau KPI dashboards.", "category": "experience", "source_location": "Resume (AI/ML and Data Engineer variants, Oct 2026)", "verification_note": "Employer-internal, no public artifact. Added to the record on 2026-10-02. This role does not appear in the LinkedIn export reviewed on 2026-09-25, which lists only Hippocratic AI, Yrall and the UMKC roles, so a reviewer comparing the two sources will find it on one and not the other. Worth adding to LinkedIn.", "project_or_experience": "Movate", "tags": [ "data-analysis", "sql", "tableau", "bi-tools", "self-reported" ] }, { "verification_status": "VERIFICATION_REQUIRED", "source_type": "self_reported", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "last_verified": "2026-10-02", "id": "skill.gcp", "claim": "Google Cloud Platform, specifically BigQuery and Cloud Storage, used for production data pipelines.", "category": "skill", "source_location": "Resume (Oct 2026); Yrall Media Solutions role", "verification_note": "Self-reported and new to the record on 2026-10-02; earlier resume variants listed no cloud platform at all. No GCP code, infrastructure definition or deployment artifact appears in any public repository, so from the public record this is adjacent at best. The demonstrated container and CI work is Docker and GitHub Actions in ClinIQ.", "project_or_experience": "Yrall Media Solutions", "tags": [ "gcp", "bigquery", "cloud", "data-engineering", "self-reported" ] }, { "id": "handshake.role", "claim": "Software Engineer Fellow at Handshake AI, remote, October 2026 to present: authors complex programming and system-design challenges that benchmark the reasoning and code-generation capabilities of leading AI models, and evaluates AI-generated code for correctness, computational efficiency and reliability, diagnosing defects, flawed reasoning and architectural weaknesses.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience", "source_type": "self_reported", "source_location": "Role description supplied by Bhavani on 2026-10-26", "verification_note": "Current role, started 2026-10-25. Self-reported. The work is confidential, so no public artifact exists and none should be expected. No figures exist yet and none are stated; any later figure must be added here as self-reported before it appears on the site.", "project_or_experience": "Handshake AI", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "code-evaluation", "llm-evaluation", "benchmarking", "system-design", "failure-analysis", "self-reported" ], "last_verified": "2026-10-26" }, { "id": "handshake.production_readiness", "claim": "Applies backend and data engineering experience (Python, FastAPI, REST APIs, SQL, PySpark, BigQuery, Airflow) to assess whether AI-generated solutions are production ready.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience", "source_type": "self_reported", "source_location": "Role description supplied by Bhavani on 2026-10-26", "verification_note": "Self-reported. The work is confidential, so no public artifact exists and none should be expected. No figures exist yet and none are stated; any later figure must be added here as self-reported before it appears on the site. This records the technologies the role draws on; it is not new evidence for any of them, and each keeps whatever standing it has elsewhere in this database.", "project_or_experience": "Handshake AI", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "backend", "python", "sql", "data-engineering", "airflow", "code-evaluation", "self-reported" ], "last_verified": "2026-10-26" }, { "id": "handshake.vqa_datasets", "claim": "Develops visual question answering (VQA) datasets for multimodal vision-language models, authoring image-grounded question-answer pairs with verified ground truth.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience", "source_type": "self_reported", "source_location": "Role description supplied by Bhavani on 2026-10-26", "verification_note": "Self-reported. The work is confidential, so no public artifact exists and none should be expected. No figures exist yet and none are stated; any later figure must be added here as self-reported before it appears on the site.", "project_or_experience": "Handshake AI", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "eval-datasets", "multimodal", "computer-vision", "annotation", "self-reported" ], "last_verified": "2026-10-26" }, { "id": "handshake.rubric_eval", "claim": "Conducts rubric-based, human-in-the-loop evaluation of model outputs across code, audio, image and text, delivering structured feedback on accuracy, hallucinations and failure modes that informs model training.", "verification_status": "VERIFICATION_REQUIRED", "category": "experience", "source_type": "self_reported", "source_location": "Role description supplied by Bhavani on 2026-10-26", "verification_note": "Self-reported. The work is confidential, so no public artifact exists and none should be expected. No figures exist yet and none are stated; any later figure must be added here as self-reported before it appears on the site.", "project_or_experience": "Handshake AI", "commit_sha": null, "exact_code_link": null, "artifact_link": null, "public_safe": true, "tags": [ "llm-evaluation", "human-in-the-loop", "annotation", "multimodal", "self-reported" ], "last_verified": "2026-10-26" }, { "id": "fundproof.provenance", "claim": "Fundproof is a two-person joint project, a two-sided lending copilot for small-business loans: the lender side turns a document packet into a cited, stress-tested recommendation that a person decides on, and the borrower side, Founder Studio, simulates cash and debt service coverage under the lender's stress test. Bhavani co-wrote the specs, build prompts and evaluation protocol and reviews the work; Claude Code implemented it. The repository on this account is her fork of the joint repository.", "verification_status": "VERIFIED", "category": "provenance", "source_type": "repository_metadata", "source_location": "README.md 'Team and this fork'; git history of the fork", "verification_note": "Up to commit da2a677 (2026-10-20), 149 of 153 commits are authored by Claude; the other 4 are three merges and one README commit from the two owners' accounts. The project's build plan sets the credit rule that each person claims only what they built, so nothing here is presented as sole work or as code Bhavani wrote by hand. The pinned commit 0efaca9 is a README rewrite made in this fork at her request on 2026-10-23 and is not yet in the joint repository.", "project_or_experience": "Fundproof", "commit_sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "exact_code_link": "https://github.com/Bhavani-abhavi/Fundproof/tree/0efaca9f871c9affd94c0fea980b522a789cdccd", "artifact_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/README.md", "public_safe": true, "tags": [ "provenance", "collaboration", "teamwork", "fundproof" ], "last_verified": "2026-10-26" }, { "id": "fundproof.limits.stated", "claim": "Fundproof states its limits: apart from models scored on public datasets (Lending Club credit, freMTPL2 insurance, a UCI German Credit fairness demo) and a few checks on public data such as metro data coverage, its evaluations run on SYNTHETIC documents, applications, worlds or founder packages, and every Fundproof figure on this site is measured on SYNTHETIC data; it is not deployed and has no users; no hosted LLM has been called, and memos come from a deterministic template.", "verification_status": "VERIFIED", "category": "research_integrity", "source_type": "code_and_artifact", "source_location": "README.md 'Synthetic data notice', 'Status', 'Results', 'Datasets and licenses' and 'Limitations'; results/agent_metrics.json llm_dependent", "verification_note": "agent_metrics.json records that no LLM call was made and that the memo writer is a deterministic template, and studio_handoff_metrics.json records the same. The README also states that most results files were written from a working tree with uncommitted changes and that several predate later code changes, and lists which. Those caveats carry over to every Fundproof claim here.", "project_or_experience": "Fundproof", "commit_sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "exact_code_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/README.md", "artifact_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/results/agent_metrics.json", "public_safe": true, "tags": [ "limitations", "synthetic-data", "integrity", "fundproof" ], "last_verified": "2026-10-26" }, { "id": "fundproof.result.feasibility", "claim": "Founder Studio's Monte Carlo feasibility simulator runs 1,000 paths over 36 months. On 2,400 SYNTHETIC well-specified worlds, where the true path is one more draw from the simulator's own model, its Brier score for P(year-2 DSCR >= 1.25) is 0.1405 against 0.1567 for a constant in-sample base rate: a narrow gain.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "code_and_artifact", "source_location": "results/studio_feasibility_metrics.json headline; src/uwcopilot/studio/feasibility/simulate.py (DEFAULT_PATHS = 1000, DEFAULT_MONTHS = 36)", "verification_note": "Committed values: simulator 0.1405 [0.1321, 0.1493], single-point forecast 0.4796 [0.4596, 0.4992], in-sample base rate 0.1567 [0.1473, 0.1664]; the simulator and base-rate intervals overlap. On misspecified worlds the simulator scores 0.1433. The worlds come from a generator written by the same team, so this checks internal consistency, not accuracy on real businesses, and the file records that no real outcome data exists. Written from a working tree with uncommitted changes (git_dirty true).", "project_or_experience": "Fundproof", "commit_sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "exact_code_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/src/uwcopilot/studio/feasibility/simulate.py", "artifact_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/results/studio_feasibility_metrics.json", "public_safe": true, "tags": [ "simulation", "monte-carlo", "calibration", "evaluation", "synthetic-data", "fundproof" ], "last_verified": "2026-10-26" }, { "id": "fundproof.result.package_traced", "claim": "36 SYNTHETIC founder packages, PDFs of 27 to 34 pages, were read back: 30,681 printed numbers, 0 untraced to the ledger. Code fills every number from the ledger, so 0 is expected by design; the check tests the plumbing, not a model.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "code_and_artifact", "source_location": "results/studio_handoff_metrics.json golden_founder_scenarios", "verification_note": "package_numbers_read_back 30681, untraced_numbers_package_text 0, untraced_numbers_pdf_text 0, pdf_pages 27 to 34. All 36 packages, and 35 of 35 SYNTHETIC startup applications, stopped at the lender side's human checkpoint with 0 decisions recorded without a human. Deterministic template memo; no language model was called. Written from a clean working tree (git_dirty false).", "project_or_experience": "Fundproof", "commit_sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "exact_code_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/scripts/eval_studio_handoff.py", "artifact_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/results/studio_handoff_metrics.json", "public_safe": true, "tags": [ "grounding", "provenance-tracking", "testing", "synthetic-data", "fundproof" ], "last_verified": "2026-10-26" }, { "id": "fundproof.agent.checkpoint", "claim": "A LangGraph agent with allow-listed tools reached the human checkpoint on 65 of 65 SYNTHETIC applications, recorded 0 decisions without a human, and sent 60 of 60 required escalations to the right queue. End-to-end routing matched the designed labels for 49 of 65 (0.7538). The memo is a deterministic template; no LLM was called.", "verification_status": "VERIFIED", "category": "evaluation", "source_type": "code_and_artifact", "source_location": "results/agent_metrics.json deterministic_path; src/uwcopilot/agent/graph.py (human_review always interrupts)", "verification_note": "reached_human_checkpoint 1.0 (n 65); decisions_recorded_without_a_human 0; all_required_escalations 1.0 (n 60); end_to_end_routing_agreement_with_designed_labels 0.7538 [0.6462, 0.8462] (n 65), which is 49 of 65. Measured on 2026-10-11 from commit 6fcdbb0 with uncommitted changes; the README records that agent files changed afterwards and the metrics were not re-run. Co-designed, implemented by Claude Code and run only on synthetic applications, so it is not production use of an agent framework (see gap.agent_frameworks).", "project_or_experience": "Fundproof", "commit_sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "exact_code_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/src/uwcopilot/agent/graph.py", "artifact_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/results/agent_metrics.json", "public_safe": true, "tags": [ "langgraph", "human-in-the-loop", "agent-reliability", "evaluation", "synthetic-data", "fundproof" ], "last_verified": "2026-10-26" }, { "id": "fundproof.mcp.server", "claim": "Fundproof's MCP server, built on the official MCP Python SDK, exposes 4 read-only tools bound to the same allow-list as the agent. Its smoke test passed 17 of 17 checks over stdio and 17 of 17 over streamable HTTP, with 0 LLM calls.", "verification_status": "VERIFIED", "category": "architecture", "source_type": "code_and_artifact", "source_location": "results/mcp_smoke_stdio.json; results/mcp_smoke_streamable_http.json; src/uwcopilot/mcp_server/server.py (READ_ONLY tool annotations)", "verification_note": "Both files: all_passed true, 17 checks, llm_calls 0, tools get_guideline_section, score_application, screen_sanctions and search_underwriting_guidelines. Run on 2026-10-11. The server has since gained a SYNTHETIC startup program (84 resources rather than 62) and the smoke test, which still expects 62, has not been re-run. This is the team project's server; the read-only MCP server in this portfolio's own repository is a separate piece of work.", "project_or_experience": "Fundproof", "commit_sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "exact_code_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/src/uwcopilot/mcp_server/server.py", "artifact_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/results/mcp_smoke_stdio.json", "public_safe": true, "tags": [ "mcp", "api-integration", "testing", "fundproof" ], "last_verified": "2026-10-26" }, { "id": "fundproof.docintel.scanned", "claim": "The document pipeline's rules extractor, which cites the page of every field, reached field F1 0.9159 on the scanned pages of 60 SYNTHETIC test packets (0.9936 overall, where most values sit on clean text-layer pages). No LLM is involved.", "verification_status": "VERIFIED", "category": "experimental_result", "source_type": "code_and_artifact", "source_location": "results/docintel_metrics.json rules.fields.by_condition.scanned", "verification_note": "Committed values: scanned F1 0.9159 [0.8736, 0.9566]; overall 0.9936 [0.9889, 0.9971]; test split of 60 packets from generator synthetic-packets-v2, generated with a new seed because the v1 test split had been inspected during review. Written from code with uncommitted changes; the README lists the clean re-run as pending.", "project_or_experience": "Fundproof", "commit_sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "exact_code_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/scripts/eval_docintel.py", "artifact_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/results/docintel_metrics.json", "public_safe": true, "tags": [ "document-ai", "ocr", "extraction", "evaluation", "synthetic-data", "fundproof" ], "last_verified": "2026-10-26" }, { "id": "fundproof.result.provenance_checker", "claim": "A number-to-source checker caught 627 of 627 planted number errors in SYNTHETIC guideline sentences with 0 of 290 false flags, and its blind spots are reported alongside: it caught only 88 of 108 comparator flips and 6 of 55 wrong-row values. Not a held-out test.", "verification_status": "VERIFIED", "category": "evaluation", "source_type": "code_and_artifact", "source_location": "results/provenance_metrics.json numeric_checker.mutation_test", "verification_note": "recall_by_type: value_perturb 179/179, magnitude_x10 184/184, unit_swap 89/89, cross_section 175/175, 627 in all; comparator_flip 88/108; wrong_row 6/55, recorded in the file as a known blind spot. False flags: 0 of 203 unmutated sentences and 0 of 87 equivalent rewrites. The checker was debugged on these reference sentences, which is why the README says it is not a held-out test. Written from a clean working tree (git_dirty false).", "project_or_experience": "Fundproof", "commit_sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "exact_code_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/src/uwcopilot/provenance/numeric_check.py", "artifact_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/results/provenance_metrics.json", "public_safe": true, "tags": [ "grounding", "testing", "evaluation", "negative-result", "fundproof" ], "last_verified": "2026-10-26" }, { "id": "fundproof.web.e2e", "claim": "The web app (FastAPI and Next.js) passed 5 of 5 Playwright end-to-end flows, 19 of 19 steps, on a local Docker container. It has not been deployed.", "verification_status": "VERIFIED", "category": "evaluation", "source_type": "code_and_artifact", "source_location": "results/web_metrics.json runs.local-container.e2e", "verification_note": "Measured on 2026-10-14 on the image of commit 271ca4d, signed in with the local dev login; Hugging Face OAuth was not exercised. The web code has changed since and has not been re-measured. A deployment workflow exists but no deploy has run, so there is no public URL.", "project_or_experience": "Fundproof", "commit_sha": "0efaca9f871c9affd94c0fea980b522a789cdccd", "exact_code_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/scripts/web_metrics.py", "artifact_link": "https://github.com/Bhavani-abhavi/Fundproof/blob/0efaca9f871c9affd94c0fea980b522a789cdccd/results/web_metrics.json", "public_safe": true, "tags": [ "testing", "e2e-testing", "web", "fundproof" ], "last_verified": "2026-10-26" } ] }