atlas-compliance-copilot
venkathub/atlas-compliance-copilot
Permission-aware, cost-routed, evaluated enterprise AI copilot for a financial/compliance domain (Spring AI + LangGraph + MCP).
Instalação
npx convez-ai venkathub/atlas-compliance-copilot1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
Atlas
Production-grade, permission-aware enterprise AI operations copilot for a financial/compliance domain.
An employee asks a question → Atlas retrieves only documents they are cleared to see → answers with citations → and, on request, executes governed actions behind a human-in-the-loop checkpoint. Every model call is cost-routed, evaluated, and traced.
Status: P0 (foundations) in progress. See docs/ROADMAP.md for the phase plan.
Architecture (6 subsystems + cross-cutting evals/observability)
| Module | Stack | Responsibility | Phase |
|---|---|---|---|
ui/ | React | Chat + admin; streamed answers, inline citations, agent trace | P5 |
gateway/ | Spring Boot | Auth, cost-aware model router, semantic cache, rate limiting | P3 |
rag-engine/ | Spring AI | Permission-aware retrieval: chunk/embed, hybrid search, citations | P1 |
| (pgvector) | Postgres + pgvector | Embeddings + role-tagged metadata (RBAC), agent checkpoints | P1/P4 |
agents/ | Python / LangGraph | Planner–executor agents, memory, human-in-the-loop | P4 |
mcp-tools/ | Spring Boot | Governed enterprise actions exposed over MCP | P4 |
evals/ | Python / RAGAS/DeepEval | Eval harness (CI-gated) + Langfuse/Grafana observability | P2 |
infra/ | Docker / CI | Compose stack, multi-arch images, GitHub Actions | P0 |
The LLM never runs locally — it lives on a remote Ollama GPU endpoint set via OLLAMA_BASE_URL
(see docs/RUNBOOK.md). All model config is env-swappable.
Quickstart (P0 — fills in as increments land)
cp .env.example .env # then edit values (never commit .env)
# docker compose up # (Increment 2) Postgres+pgvector + Redis
# make health # (Increment 2) assert services healthy
# mvn -pl rag-engine test # (Increment 3) unit tests; live smoke test is @Tag("live")
Documentation
docs/ROADMAP.md— vision, phase plan P0–P5, risks, skills map, security baselinedocs/DECISIONS.md— architectural decision log (ADRs)docs/RUNBOOK.md— local host + cloud Ollama (JarvisLabs) operationsdocs/PORTFOLIO.md— quantified, resume-ready outcomes per phaseCLAUDE.md— engineering operating agreement
License
Apache-2.0 © 2026 Venkatesh.
Note: the FinanceBench dataset used from P1 stays under its own CC-BY-NC-4.0 license
(see docs/DECISIONS.md ADR-0004) — that covers data, not this code.