docs(samples): add AML alert triage coded agent sample

[1aifanatic]

Closes #1109

Summary

Adds samples/aml-alert-triage-agent: a LangGraph coded agent that triages anti-money-laundering alerts. The UiPath LLM Gateway does the language work (extraction and rationale), and an external decision model (TypeSafe Jev) makes every typed, calibrated decision in a single call.

extract (LLM Gateway) -> decide (Jev: 5x red flag, risk score, disposition) -> explain (LLM Gateway)

What's in the sample

File	Purpose
main.py	The graph: extract -> decide -> explain, with a switchable decider input (jev / llm)
rubric.py	The policy: red flags, risk levels, and dispositions, versioned via RUBRIC_VERSION
evaluate.py	Schema gate, accuracy, evidence grounding, and a --compare head-to-head
evals/	12 synthetic alerts with ground truth, plus ready-to-run sample-*.json inputs
docs/	Architecture SVG, Mermaid diagrams, copy-paste test inputs

The layout matches the other samples: langgraph.json, uipath.json, bindings.json, .env.example, and a uv-style pyproject.toml with a uipath-dev dev group. entry-points.json is not committed, per the repo's .gitignore. samples/README.md gets a new entry.

Notes for reviewers

• Third-party dependency: the default decider calls the TypeSafe API (typesafe-sdk) and needs a JEV_API_KEY. Locally the key comes from .env; in serverless runs it is read from an Orchestrator Asset named JevApiKey. Setting decider: "llm" runs the whole agent on the LLM Gateway alone.
• Tested: it runs locally with uipath run, and was published and run on Orchestrator as PythonCodedAgent. After the review fixes (902ec88), the raw narrative is no longer sent to Jev, so the README's measured results are marked pending re-measurement until python evaluate.py --compare is re-run. The review fixes are covered by an offline check that runs the real graph with both models faked: prompt-injection containment, input-field protection, output bounds, and the compare exit code.
• Untrusted input: alert text goes to every LLM call in its own <alert>-tagged message, and all model output is validated against Pydantic schemas (ExtractedFacts, LLMDecision, bounded Output).
• CI: the testcases/* integration jobs and SonarCloud fail on this fork PR because repository secrets (CLIENT_ID/CLIENT_SECRET/BASE_URL, SONAR_TOKEN) are not exposed to forks. They fail at uipath auth and don't exercise this sample. Lint, unit tests, commit-lint and Socket Security pass.
• All alert data is synthetic.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G2Vpmcrr6qxtrpnwynoJJk

[Copilot]

Copilot review overview

🟡 Changes recommended

Critical data-handling and validation issues, along with evaluation and dependency fixes, remain unresolved.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 3 · 2

Open (5)

• Protect extraction prompts from narrative prompt injection · New
• Omit raw narratives from the external Jev digest · New
• Prevent unvalidated facts from overwriting authoritative inputs · New
• Require both model results to pass schema validation · New
• Enforce probability and confidence bounds · New

What changed in this PR

Adds an AML alert-triage LangGraph coded-agent sample integrating UiPath LLM Gateway with TypeSafe Jev.

Changes:

• Adds extraction, decision, and explanation graph nodes with switchable deciders.
• Adds AML rubric, synthetic evaluations, and comparison tooling.
• Adds deployment configuration, examples, diagrams, and documentation.

File	Description
samples/​README.md	Adds the sample index entry.
samples/​aml-alert-triage-agent/​uipath.json	Configures packaging and runtime.
samples/​aml-alert-triage-agent/​rubric.py	Defines the AML decision rubric.
samples/​aml-alert-triage-agent/​README.md	Documents usage, architecture, and results.
samples/​aml-alert-triage-agent/​pyproject.toml	Defines project dependencies.
samples/​aml-alert-triage-agent/​main.py	Implements the agent graph and integrations.
samples/​aml-alert-triage-agent/​langgraph.json	Registers the graph entry point.
samples/​aml-alert-triage-agent/​evaluate.py	Runs evaluation and comparison reports.
samples/​aml-alert-triage-agent/​evals/​sample-llm.json	Provides a runnable LLM example.
samples/​aml-alert-triage-agent/​evals/​sample-escalate.json	Provides a runnable escalation example.
samples/​aml-alert-triage-agent/​evals/​sample-close.json	Provides a runnable closure example.
samples/​aml-alert-triage-agent/​evals/​sample-ambiguous.json	Provides a runnable ambiguous-alert example.
samples/​aml-alert-triage-agent/​evals/​alerts.json	Provides synthetic alerts and expected outcomes.
samples/​aml-alert-triage-agent/​docs/​test-examples.md	Documents testing examples.
samples/​aml-alert-triage-agent/​docs/​diagrams.md	Documents system flows.
samples/​aml-alert-triage-agent/​docs/​architecture.svg	Provides the architecture illustration.
samples/​aml-alert-triage-agent/​bindings.json	Defines deployment resources.
samples/​aml-alert-triage-agent/​agent.mermaid	Records the graph topology.
samples/​aml-alert-triage-agent/​.gitignore	Ignores generated evaluation results.
samples/​aml-alert-triage-agent/​.env.example	Documents required environment variables.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.