In active development
ClearFlag
A transaction fraud rules engine with explainable, agent-written rationales.
ClearFlag scores transactions against configurable fraud rules and shows analysts exactly which rules fired and why. Its transparency layer exposes every rule hit, and the Investigation Agent now being built uses LangGraph to gather evidence through three tools (transaction history, merchant risk, geographic distance) before writing a rationale.
The point is the same one an auditor makes: a flag nobody can explain is a flag nobody can defend.
Decisions I'd walk you through
- Deterministic tool dispatch. Code decides which tools run, not the LLM, so every investigation follows a repeatable, testable path.
- Append-only rationale table. Rationales are composed outside the request path and stored as both a cache and an audit record.
- Graceful fallback. If the agent, a tool, or a guardrail fails, the system falls back to a template formatter instead of showing nothing.
- Python
- FastAPI
- React
- LangGraph
- SQLAlchemy
- Alembic
- Neon Postgres
- GitHub Actions
- LangSmith