jev-uipath-coded-agent
UiPath coded agent for AML alert triage where an LLM reads and writes but Jev makes every red-flag, risk and disposition decision.
Overview
This financial-services demo is a UiPath coded agent that triages anti-money-laundering alerts, splitting work along a clear seam. The UiPath LLM Gateway (gpt-5-mini) reads the alert, extracts parties and amounts, does all arithmetic and afterwards writes the rationale. TypeSafe's Jev makes every decision in one call with seven parallel questions: five red-flag Nouls (structuring, rapid pass-through, shell company indicators, high-risk jurisdiction, purpose mismatch), a risk-level Score from Low to Critical, and a disposition Choice of escalate, close or need_info.
The agent is built with LangGraph, packaged with the UiPath CLI and runs in UiPath Automation Cloud's serverless runtime, with the Jev key stored as an Orchestrator asset. A 12-alert synthetic evaluation with hand-authored ground truth compares Jev with the gateway LLM answering the same seven questions: both scored 10 of 12, with a median decision of 398 ms for Jev versus 6,160 ms and a set cost of $0.000426 versus 2.4 platform units. The README explicitly says Jev wins on speed and cost, not on better answers, and that n=12 is not statistically meaningful.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this alert show structuring / rapid pass-through / shell company indicators / high-risk jurisdiction / purpose mismatch? | — |
| Score | What is the risk level of this alert? | Low · Medium · High · Critical |
| Choice | What should the disposition be? | escalate · close · need_info |
How it works
- LLM extracts parties and amounts and does arithmetic
- One Jev call answers seven questions on the shared state
- Disposition and risk drive the agent's action
- LLM writes the rationale from Jev's decisions
Strengths
- README reports median 398 ms per decision versus 6,160 ms for the LLM
- Tiny cost for the evaluation set
- Clean separation of language work and decisions
- Runs inside UiPath's governed runtime
Caveats
- 12 synthetic alerts; accuracy comparison not meaningful
- Demo, not a compliance-validated system
- Requires UiPath Automation Cloud
Getting started
python evaluate.py --compare, or `uipath run agent -f evals/sample-escalate.json` for one alert.At a glance
- Type
- App
- Status
- prototype
- Language
- Python
- Runtime
- UiPath coded agent (LangGraph)
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
FINS demo: a UiPath coded agent for AML alert triage where every decision is…
wh000wh000/awesome-jev-live
Frequently asked questions
Is Jev more accurate than the LLM here?
No. Both scored 10 of 12; the README says Jev wins on speed and cost only.
What does the LLM still do?
Reads the alert, extracts entities, does arithmetic and writes the rationale.
Where does it run?
In UiPath Automation Cloud's serverless runtime, with the Jev key as an Orchestrator asset.
Related projects
Crypto exchange risk gateway where Jev triages deposits and withdrawals while hard rules and Java code make the final call.
ServiceprototypeJavaSAP S/4HANA proof of concept that sends anonymised purchase-order risk indicators to Jev and classifies orders to release, approve or block.
ProjectprototypeABAPWorkflow design article: one Jev Noul asks whether a domain name refers to a protected brand, with the TLD kept out of the input.
ArticleproductionToolkit for Jev decision contracts: compiled policies, egress controls, tracing, stability checks and certified decision SLOs.
ProjectbetaTypeScriptLocal-first personal-data inventory tool with optional Jev classification per column, not per row, and dry-run payload review.
ProjectprototypePythonSpanish compliance triage: clauses extracted deterministically, then Jev labels each obligation, risk or no action.
LibrarybetaPython