12 October 2026 client-inclusive measurement: 4,800 generated unlabeled scores over 120 seconds, four clients at 10 requests/s each, zero failures. Not saturation or production latency.
Evaluation and serving can disagree when features include future information or retries recalculate accepted transactions.
The implementation
What I built
Shared prior-only feature construction between offline and serving paths, then added atomic SQLite replay and role-controlled inspection.
FastAPI
SQLite
scikit-learn
Python
JavaScript
Features
Inspect original saved scoring results
Retry accepted transactions without re-scoring
Use tenant-scoped reader, writer, and admin roles
Architecture
One feature module for offline preparation and serving.
Versioned scikit-learn artifact, runtime manifests, and export/reload parity checks.
Tenant-scoped FastAPI, SQLite replay history, and browser workbench.
Engineering decisions
Exclude the current observation from its own historical counts and amount means.
Commit input, original features, response, revision, and audit together.
Return the original saved result on exact retries.
Validation and measured results
The 10 October 2026 historical study used a chronological 100,000-row prefix with separate fit, calibration, selection, and historical final partitions.
Historical fixed-threshold precision was 9.23% and recall was 41.28%; most alerts were false and most fraud was missed.
The dated hosted engineering baseline at 676a4795ba04a7bb2acf2bf573826a6e9d77311a, run 38409317955, passed 404 Python tests with zero failures, errors or skips. CI uses synthetic model fixtures and separately exercises parity, roles, retries, browser behavior, restart recovery, and same-host restore.
A separate 12 October 2026 local workload completed 4,800 generated unlabeled scores over 120 seconds, with four clients paced at 10 requests/s each, zero failures and client-inclusive p95 16.77 ms. This is one selected-model development-host measurement, not fraud accuracy, saturation, a speedup, or a production SLA. Saved receipt: reports/operating-2026-10-12/selected-model-001/report.json; documented in docs/resume-contract.md.
The repository was not publicly accessible on 14 October 2026. These notes summarize locally inspected source and receipts; downloadable receipts are not attached.
Limits and unfinished work
Historical full-source exposure cannot be ruled out. The old final partition is development-only for future work.
No independently unseen production fraud-prevention result is claimed.
Raw transactions and model weights remain unpublished; redistribution permission is unreviewed.
10 October historical quality study, 12 October local operating workload and dated hosted engineering baseline remain separate. The repository and CI are not publicly accessible; receipt filenames and scope are listed above.
