AassemD/acme-bank-concierge

Acme Bank Concierge — demo agent for Vidimus

A multi-agent banking customer-service assistant, built on Google ADK with MCP tools and DeepSeek-V3.1 (open-source, via OpenRouter's free tier) as the underlying LLM. Designed as the target for Vidimus's end-to-end test harness — every "planted weakness" below is something Vidimus's probes or evidence flow is expected to catch.

The LLM is swappable via the LLM_MODEL env var — see agents/llm.py for free-tier alternatives (Qwen3, Llama 3.3, DeepSeek-R1).

Architecture

        ┌─────────────────────────────┐
        │  POST /chat   (FastAPI)     │
        └───────────────┬─────────────┘
                        │
              ┌─────────▼─────────┐
              │ account_concierge │  root LlmAgent
              │   (routes intent) │
              └─────┬───┬───┬───┬─┘
                    │   │   │   │
   ┌────────────────┘   │   │   └──────────────────┐
   │              ┌─────┘   └─────┐                │
   ▼              ▼               ▼                ▼
balance_      disputes_       refunds_         escalation_
specialist    specialist      specialist       specialist

   │              │               │                │
   └──── MCP toolset (stdio) ─────┴────────────────┘
                    │
        ┌───────────▼────────────┐
        │  mcp_server/server.py   │
        │  banking tool surface   │
        └─────────────────────────┘

• —Root (account_concierge) — greets, identifies intent, delegates.
• —Sub-agents — each owns one workflow and only the tools it needs.
• —MCP server — single stdio process exposing the bank's tools. Spawned by ADK's MCPToolset on first call; lives for the container's lifetime.

Local dev

cd demo-agent
python -m venv .venv
. .venv/Scripts/activate     # PowerShell: .\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
cp .env.example .env         # then fill in OPENROUTER_API_KEY
python app.py

Then in another shell:

curl -s -X POST http://localhost:8080/chat \
  -H "content-type: application/json" \
  -d '{"messages":[{"role":"user","content":"hi, what is my balance? my account is AC-100001"}]}' | jq

Deploy to Cloud Run (Paris / Brussels)

export OPENROUTER_API_KEY=...           # from https://openrouter.ai/keys (free)
export DEMO_AGENT_TOKEN=$(openssl rand -hex 16)
gcloud config set project <your-gcp-project>
./deploy_cloud_run.sh

The script prints the public Cloud Run URL on success. Region defaults to europe-west1 (Belgium); override with REGION=europe-west9 for Paris.

Vidimus agent-intake config

After deploy, paste the following into Vidimus's New Agent form so its test harness can call this service:

Vidimus's classifier should mark this customer-facing + financial-impact

• —critical, which will pull in Art 5, 9, 12, 13, 14, 15, 16 obligations.

Planted weaknesses (Vidimus is expected to find these)

The risk_management and documentation gaps are evidence_required in Vidimus — they can't be verified by chat probes alone. To clear those, upload a (fake) risk-management policy document via Vidimus's evidence-upload flow; the document verifier will grade it and the proof report should flip the obligations from needs-action to document-verified.

Updating the demo

• —New planted weakness → edit the sub-agent's instruction string + add a row to the table above.
• —New tool → add it to mcp_server/server.py, wire it into the agent that should own it via tools=[toolset_for(...)].
• —Swap the LLM → set LLM_MODEL env var to any LiteLLM-supported model id (e.g. openrouter/qwen/qwen3-235b-a22b:free, openrouter/meta-llama/llama-3.3-70b-instruct:free). No code change.