An AI-powered NBA take analyzer that fact-checks basketball opinions with real stats, generates visual stat cards, and roasts bad takes on X (Twitter).
"@LeGMLab LeBron is washed"
"Bro averaged 25.7/7.3/8.3 last season on 54% TS at age 40. Your take is washed, not LeBron. Respectfully."
@mention or search tweet
-> LLM analyzes the take (Claude / OpenAI)
-> Fetches real NBA stats via tools (season avgs, game logs, advanced stats)
-> Generates verdict (trash / mid / valid) + roast (<280 chars)
-> Renders stat comparison chart (optional)
-> Posts reply with chart attached
The agent runs a multi-turn tool-use loop: the LLM decides which stats to look up, gets the results, then writes its verdict backed by real numbers.
- Take analysis -- Verdicts (trash/mid/valid) with confidence scores and stat-backed reasoning
- 6 stat tools -- Season averages, recent games, advanced stats, player comparisons, team standings, team records
- Chart generation -- Dark-themed stat cards and comparison tables with winner indicators
- X bot -- Reactive (replies to @mentions) with rate limiting and monthly budgets
- REST API -- FastAPI endpoints for programmatic take analysis
- Multi-LLM -- Claude, OpenAI, or any OpenAI-compatible provider
| Stat Card | Flexible Comparison |
|---|---|
 |
 |
- Python 3.12+
- uv package manager
git clone https://github.com/KenWuqianghao/LeGM-Lab.git
cd LeGM-Lab
uv syncecho 'ANTHROPIC_API_KEY=sk-...' > .env
uv run python scripts/run_single_take.py "Steph Curry is the greatest shooter ever"uv run uvicorn legm.main:app --reloadPOST /api/v1/takes/analyze {"take": "Jokic is the best center ever"}
GET /api/v1/takes # list recent analyses
GET /health # liveness probe
# Add Twitter creds to .env (see Configuration below)
BOT_DRY_RUN=true uv run python scripts/run_bot.pyAll config is via environment variables (or .env file):
| Variable | Description |
|---|---|
ANTHROPIC_API_KEY |
Claude API key |
LLM_PROVIDER |
claude, openai, or openai_compat |
LLM_MODEL |
Model name (default: claude-sonnet-4-6) |
OPENAI_COMPAT_BASE_URL |
Base URL for OpenAI-compatible endpoint |
OPENAI_COMPAT_API_KEY |
API key for OpenAI-compatible endpoint |
OPENAI_COMPAT_MODEL |
Model name for OpenAI-compatible endpoint |
DATABASE_URL |
SQLAlchemy URL (default: SQLite) |
TWITTER_API_KEY |
X API key |
TWITTER_API_SECRET |
X API secret |
TWITTER_BEARER_TOKEN |
X bearer token |
TWITTER_ACCESS_TOKEN |
X access token |
TWITTER_ACCESS_TOKEN_SECRET |
X access token secret |
TWITTER_BOT_USER_ID |
Bot's numeric X user ID |
BOT_DRY_RUN |
true to log without posting |
BOT_PROACTIVE_ENABLED |
true to enable proactive search loop |
BOT_MENTION_POLL_INTERVAL |
Seconds between mention polls (default: 60) |
BOT_MONTHLY_BUDGET |
Max tweets per month (default: 450) |
The bot runs on Railway as a background worker. A Dockerfile and railway.toml are included.
# Local testing
docker compose up --build
# Railway
# 1. Connect GitHub repo in Railway dashboard
# 2. Add a volume mounted at /data
# 3. Set env vars in Railway UI
# 4. DeployThe custom LeLM model (Qwen3-8B fine-tuned on NBA takes) is served via Modal on a T4 GPU with scale-to-zero. It exposes an OpenAI-compatible /v1/chat/completions endpoint that the bot connects to via the openai_compat provider.
# Deploy
modal deploy deploy/modal_lelm.py
# Test
curl -X POST https://<modal-url>/v1/chat/completions \
-H "Authorization: Bearer <AUTH_TOKEN>" \
-H "Content-Type: application/json" \
-d '{"model":"lelm","messages":[{"role":"user","content":"Is LeBron washed?"}]}'Bot env vars to connect:
LLM_PROVIDER=openai_compat
OPENAI_COMPAT_BASE_URL=https://<modal-url>/v1
OPENAI_COMPAT_API_KEY=<AUTH_TOKEN>
OPENAI_COMPAT_MODEL=lelm
src/legm/
agent/ # LLM agent loop, system prompt, tool definitions
api/ # FastAPI routes (takes, bot control, health)
db/ # SQLAlchemy models, migrations, repository
llm/ # Multi-provider LLM abstraction (Claude, OpenAI, LeLM)
stats/ # NBA API client, caching, chart generation
twitter/ # X bot, tweet service, rate limiter, filters
config.py # Pydantic settings
main.py # FastAPI app
deploy/
modal_lelm.py # Modal deployment for LeLM inference
scripts/
run_bot.py # Bot CLI entrypoint
run_single_take.py # Analyze one take from CLI
run_live_takes.py # Batch analysis demo
uv run pytest -x -v # run tests
ruff check --fix . && ruff format . # lint + format
uv run python -m mypy . # type check- Backend: Python 3.12 / FastAPI / SQLAlchemy / Pydantic v2
- AI: Claude (Anthropic) / OpenAI / LeLM with tool-use agent loop
- Data: NBA API (nba_api) / Matplotlib for chart rendering
- Bot: Tweepy (X API v2) / asyncio event loops
- Infra: Docker / Railway / Modal (GPU inference) / SQLite (dev) / PostgreSQL (prod)
MIT