Home / Projects / Transaction Monitoring Alert Triage
Transaction Monitoring Alert Triage
Transaction Monitoring Alert Triage is a project by Akash Renuka Ashok (AI Engineer (independent prototype)). An AML alert triage pipeline prototype: deterministic typology rules raise alerts, a scikit-learn layer ranks them, and a mandatory analyst disposition gate records every decision.
- Role
- AI Engineer (independent prototype)
- Category
- Financial Crime Prototype
- Stack
- Python, pandas, scikit-learn, FastAPI, SQLite, React, Vite, Ollama
The problem
Transaction monitoring raises more alerts than analysts can review, and a model that ranks or explains them is only usable if every decision stays with a human and can be audited.
What Akash owned
Designed and built the pipeline solo: typology rules, ranking layer, analyst review interface, audit trail and the LLM narrative layer.
Highlights
- Writes every analyst disposition to an append-only audit trail behind a mandatory review gate
- Confined the local LLM to analyst narratives only, enforced by an adversarial boundary test and an import-graph assertion; if the LLM is unavailable the alert queue is unchanged
- Audited the PaySim benchmark before trusting results on it and judged it unsuitable for behavioural monitoring, consistent with its published label-leakage limitation, then built a synthetic data generator as an alternative
- Reports PR AUC, precision at k and false negatives rather than accuracy, with an automated warning when a model clears 0.99 PR AUC
Outcome
Demonstrated an auditable triage design where rules and a ranking model prioritise the work, an analyst makes every decision, and the LLM only writes.
More work by Akash Renuka Ashok:
Last updated