CMO Copilot

An AI copilot for marketing budgets — a 6-agent Qwen society whose own mistakes a learned memory gates, over a realistic ad account (~300 campaigns, $1.2M/month). Built for the Qwen Cloud hackathon on Qwen (Alibaba Cloud Model Studio) with a Model Context Protocol server.

Headline result: 27% → 100% on a 100-question CMO benchmark generated deterministically in Python (no model in the generation; every correct answer is proven present in the data before it counts).

The hard part isn't a smart model — it's reliability on decisions where the obvious move is a trap (a conversion drop that's a broken pixel, not a real decline). A planner behind a risk/trap gate solves every tier; a memory that learns which of the society's own calls to override — a decision tree fit on its own outcome history, kept only when history proves it mattered, enforced deterministically — gets there without the gates being hand-coded.

Layout

cmo/       the engine: config, datagen, scenarios, portfolio, modeling,
           tools, policy, llm (Qwen client), agents, harness, benchmark,
           multi_item, build_review, mcp_server
tracks/    track1 (learned-gate memory) · track3 (agent society) · track4 (autopilot)
api/       FastAPI backend (landing page + JSON API) — the deployable surface
scripts/   check_live.py, export_data.py (standalone utilities)
docs/      SUBMISSION · ARCHITECTURE · DEPLOY · DEVPOST · DEMO
web/        optional Next.js dev UI (not required to run or deploy)
tests/     170 offline tests

Related MCP server: flour-ads

Quickstart (offline, no key)

pip install -r requirements.txt
pytest -q                                             # 170 tests
python -m cmo.harness --agent mock                    # canary = 4.8/11
python -m cmo.benchmark --mock                        # the 100-question benchmark
python -m tracks.track1.memory_gates --sessions 4 --base society   # memory -> 100%
python -m cmo.build_review                            # -> benchmark_review.html
python -m cmo.mcp_server                              # the MCP server (stdio)

Live on Qwen Cloud

Put your key in .env (DASHSCOPE_API_KEY=..., LLM_PROVIDER=dashscope — see .env.example), then:

python scripts/check_live.py        # one-shot connectivity check
python -m cmo.bench_live            # the 7-approach comparison, live on Qwen

Endpoint defaults to the international Model Studio compatible-mode URL; set QWEN_BASE_URL for CN-region accounts. Model ids default to qwen-plus / qwen-flash / qwen-max — verify the exact strings in your console.

Deploy

docs/DEPLOY.md — a container on Alibaba Cloud (ECS / Function Compute / SAE). docker build -t cmo-copilot . && docker run -p 8000:8000 -e DASHSCOPE_API_KEY=sk-... cmo-copilot, then open / for the landing page and /api/health for status.

The story

docs/SUBMISSION.md is the full write-up (the seven-architecture experiment, the honest negative results, and the learned-gate memory). docs/ARCHITECTURE.md has the system topology and the society+memory loop. MIT licensed.