SAR investigation platform

The problem

Investigators received a case file, or pulled data from several systems and stitched it together by hand.
They checked each red flag, gathered evidence for the ones that fired, drafted a narrative, and revised it until a supervisor approved it.
A typical case took 6 to 8 hours. Even with a ready case file, checking every red flag, marking the evidence, and writing the narrative took most of that time.
The goal was to automate the manual steps and keep the investigator and supervisor as the people who decide what gets filed.

Auditability. A SAR is a regulatory filing, so every conclusion had to be traceable months later.
Data residency. Case data could not leave the bank, so the platform deploys on client infrastructure and calls models inside the client's own cloud account.

Deterministic wherever possible. Checks give the same result on every run, and the fixed flow from red flags to narrative runs as plain code rather than an agent choosing each step.
LLMs only where judgment is needed. Each red flag gets a scoped agent that runs checks as tools and calls the model only for the parts of a rule that need interpretation, rather than sending the full rules and case file to a model. This cuts cost and limits the room for error.
Configurable rules. FinCEN red flags are deliberately general, and fixed thresholds would show bad actors where to stay under the line. Rules are ingested as checks each bank configures with its own thresholds, which also makes the platform reusable across institutions.
Verification before every filing. An LLM judge scores each finding against its evidence and sends low-confidence ones to people, and an investigator and supervisor approve every case before filing.
A full audit trail. Every output is stored in write-once storage, with the rule and prompt versions that produced it.

Data scientists with frontend and backend engineers.
I designed the architecture, built the MVP, guided the team through the build, and solved the technical problems along the way.

About 50% less investigation time against the 6 to 8 hour manual baseline, in internal benchmarks during my time on the project.
Investigator and supervisor review flow unchanged.
New institutions are onboarded by loading their rules and thresholds, not by rebuilding the platform.

Bring experienced SAR investigators in from the start to validate results on real cases before building anything sophisticated around them.