Agentic Risk Desk

A transaction-analyst agent, my LangGraph learning build, in progress.

A transaction-analyst agent that cites the rules it applied and stops for a human on high-risk cases. I'm building it to learn LangGraph properly. Stage 1 of 4 is done; the rest is planned, and this page says which is which.

01 · Progress
stage 1 of 4
rules service done; the agent is next
02 · Detection
5/5
planted patterns caught on synthetic data, 0 false alarms
03 · Tests
59
tests passing on the rules service
01 · Problem

An analyst that shows its rules, and knows when to stop

A useful risk agent has to do two things a plain chatbot won’t: cite the exact rules behind every flag, and hand the high-risk cases to a person instead of deciding them itself. I’m building one to learn LangGraph on a problem where those two constraints matter.

Learning buildEverything here runs on synthetic data. It is a learning project, not a product, and nothing runs in production.

02 · The plan

Four stages, one done

fig.1 · risk-desk · plan stage 1 of 4
FIG. 1 · Green = done · dashed = planned. Stage 2 adds the LangGraph analyst and the human approval step.
  1. S1: Rules service (done)

    Deterministic rules over transactions with sliding windows and categorical escalation, behind an assess endpoint.

  2. S2: LangGraph analyst (planned)

    An agent that reads the rule hits, writes a structured decision memo, and pauses for human approval on high-risk cases.

  3. S3: API and observability (planned)

    An async API with streaming, tracing, latency and cost per case, packaged in Docker.

  4. S4: Cloud deploy (planned)

    A deploy to AWS Bedrock AgentCore, plus a comparison when the model is swapped.

03 · Built so far

Stage 1: the rules service

The rules service applies 6 AML-style rules with sliding windows and categorical escalation, and exposes single and bulk assessment endpoints. It is covered by 59 tests.

I tested it on a synthetic set of 303 transactions from 20 customers with suspicious patterns planted in it: it caught 5/5 of them with no false alarms. The same run surfaced a real bug in my design: the thresholds could never reach the HIGH level, so I recalibrated them.

04 · Key numbers

Stage 1, on synthetic data

Planted patterns
5/5

Caught, with 0 false alarms.

Tests
59

On the rules service.

05 · What's next

The agent itself, then the cloud

Next is the LangGraph analyst with human approval, then the API with tracing and cost per case, then the AgentCore deploy, planned through November 2026. This page will say what is done as each stage lands, with its own measured number: memo accuracy, latency, cost per case.