Explainable trade-surveillance platform. Synthetic FIX 4.4 order/execution flow streams through a governed lakehouse medallion; stateful Scala detectors flag market-abuse patterns (spoofing/layering, wash trading, front-running, momentum ignition, marking-the-close); a calibrated, SHAP-explained model scores each alert; and an AI agent drafts a regulation-cited Suspicious Trading Report (STR), gated by a closed-world citation check and a CI hallucination gate. Every step is recorded in a tamper-evident hash chain.
Paste a FIX message, watch it get classified, see the SHAP waterfall, the violated UMIR/MiFID rule, and the auto-drafted STR on one screen, backed by a verifiable audit badge.
Regulated trade surveillance needs AI that is useful, explainable, and auditable. SentinelTrace keeps a deterministic rule score authoritative (reproducible with no model in the loop), reports ML score + SHAP + conformal confidence alongside it, grounds every regulatory citation in a governed corpus, and makes every decision replayable through a single audit chain.
FIX generator -> Kafka/Event Hubs -> medallion (bronze/silver/gold)
-> Scala stateful detectors -> canonical Alert
-> ML scoring + SHAP -> governance (lineage, RLS, PII masking)
-> AI agent drafts regulation-cited STR -> eval gate (CI)
-> web console (one-screen demo)
| Path | Purpose |
|---|---|
generator/ |
Synthetic FIX 4.4 generator with labeled abuse scenarios |
pipeline/ |
Medallion transforms + the canonical contract + shared audit chain |
scala-detectors/ |
Stateful market-abuse detectors (Scala, sbt) |
ml/ |
Scoring, calibration, SHAP, conformal prediction |
agent/ |
STR drafting agent (tools, grounding, safety) |
evals/ |
Eval harness + quality gate |
web/ |
Next.js surveillance console |
infra/ |
Terraform + Bicep + Databricks Asset Bundles |
tests/ |
Cross-language contract + integration tests |
Early development. Built local-first (Docker) before any cloud deployment.
MIT