A modular, model-agnostic platform for enterprise AI copilots. Each module is an independently deployable, evaluated AI system sharing common platform services (auth, AI orchestration, retrieval, evaluation, observability, deployment).
🔴 Live demo: https://copilot-platform.onrender.com
Try it: open the URL for the Diagnose console + Eval Dashboard, or hit the API directly:
curl https://copilot-platform.onrender.com/health
curl -X POST https://copilot-platform.onrender.com/modules/erp-sync/agent-diagnose \
-H 'content-type: application/json' \
-d '{"query":"stripe payout does not match deposit","inputs":{"gross_payments":[600,400],"processor_fees":0,"bank_deposit":970}}'(Free-tier host — first request after idle may take ~30–50s to wake.)
Enterprise teams adopt AI fastest when copilots are trustworthy, evaluated, and deployable into real workflows. This platform demonstrates that pattern, starting with a module drawn from a real, high-friction enterprise problem: ERP sync & reconciliation failures.
| Module | Status | What it does |
|---|---|---|
| ERP Sync Reconciliation Copilot | ✅ Phases 1–5 complete | Diagnoses why an ERP/accounting sync or reconciliation failed; explains root cause with citations; suggests a fix. Every answer is scored by the enforcing eval framework. |
| Platform core | ✅ Phase 1 | Model-agnostic provider router + FastAPI gateway (/health, /providers, /chat) + Docker + Postgres/pgvector. |
| Retrieval (RAG) | ✅ Phase 2 | Pluggable embedder + vector store; /modules/erp-sync/diagnose grounds answers in a knowledge base with citations. |
| Diagnostic agent + MCP | ✅ Phase 3 | Multi-step agent (plan→act→synthesize) combining RAG + computational tools (reconciliation calculator, mapping validator); tools exposed over MCP. /modules/erp-sync/agent-diagnose. |
| Evaluation framework | ✅ Phase 4 | Golden dataset + metrics (retrieval/tool/groundedness) with an enforcing CI eval-gate — quality regressions block merges. |
| Integration Health Monitor | planned | Detects sync drift, generates plain-language incident explanations |
| Customer Onboarding Copilot | planned | Guides + live-validates a customer's integration setup |
| AI Support Engineer | planned | Ticket triage, deflection, runbook answers |
| AP Invoice Intelligence | planned | Extracts/validates invoices, flags anomalies, routes approvals |
| Evaluation Dashboard | planned | Visualizes eval scores across modules over time |
| Admin Console | planned | Tenant/config/observability administration |
Auth (OIDC/JWT) · AI Orchestration (multi-agent) · Retrieval (RAG) · Evaluation framework · Model-agnostic provider router (Anthropic / OpenAI / Bedrock / Vertex) · MCP server · Observability (OpenTelemetry) · Persistence (Postgres + pgvector).
Run the API (no API keys needed — uses the built-in mock provider):
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
uvicorn core.api.main:app --reload
# open http://localhost:8000/docsWith Docker + Postgres/pgvector:
cp .env.example .env # optionally add a provider key
docker compose up # api on :8000, db on :5432Try it:
curl localhost:8000/health
curl localhost:8000/providers
curl -X POST localhost:8000/chat -H 'content-type: application/json' \
-d '{"messages":[{"role":"user","content":"Why did my QBO deposit fail to reconcile?"}]}'Set ANTHROPIC_API_KEY or OPENAI_API_KEY in .env and pass "provider":"anthropic" (or "openai") to use a real model — the API is identical (see ADR-002).
See ARCHITECTURE.md and docs/adr/ for decision records.
- If it isn't evaluated, it isn't shipped — every module ships with an eval suite gated in CI.
- Model-agnostic — providers are swapped by config, never by code change.
- Modular — modules depend on platform services, never on each other.
- Every phase ships a public artifact — see docs/BUILD_PRINCIPLES.md.
MIT