Real Time - Trade Eligibility Classifier

RT-TEC is a real-time trading decision system that uses machine learning models to determine trade eligibility in sub-10ms latency.

Tech Stack: Python 3.11 • FastAPI • NVIDIA Triton • Feast • Redis • PostgreSQL • RedPanda • Prometheus • Grafana • Docker

Trade Eligibility Classifier — v1.3

• Feast online feature (Redis) with dev TTL.
• /debug/features and /debug/triton endpoints.
• Prometheus + example Grafana dashboard.
• Golden test scaffolding and k6 smoke.

Quick start

# Build API image (Feast + deps)
docker compose build --no-cache api

# Export model to Triton repo
docker compose run --rm -w /app api python scripts/export_dummy_model.py

# Generate Feast data, apply & materialize
docker compose run --rm -w /app/feast_repo api python data/generate_microstructure.py
docker compose run --rm -w /app/feast_repo api feast apply
docker compose run --rm -w /app/feast_repo api feast materialize-incremental "$(date -u +"%Y-%m-%dT%H:%M:%S")"

# Start core services
docker compose up -d redis postgres triton api

# Sanity
curl -s http://localhost:8080/healthz
curl -s "http://localhost:8080/debug/features?symbol=BTC"
curl -s "http://localhost:8080/debug/triton?n=8"
curl -s -X POST http://localhost:8080/v1/score -H 'Content-Type: application/json'   -d '{"symbol":"BTC","ts_ns":1,"features":[0.1,0.2,0.0,0.3,0.1,0.0,0.2,0.1],"freshness_ms":10}'

Steps:

# build & start new services
docker compose build ingest
docker compose up -d redpanda redpanda-console ingest

# export primary (v1) and canary (v2) models
docker compose run --rm -w /app api python scripts/export_dummy_model.py
docker compose restart triton

# (optional) keep Feast materialized too
docker compose run --rm -w /app/feast_repo api feast apply
docker compose run --rm -w /app/feast_repo api feast materialize-incremental "$(date -u +"%Y-%m-%dT%H:%M:%S")"

# enable canary and restart API
export CANARY_ENABLED=true && export CANARY_VERSION=2
# or set in docker-compose.yml env for api before bringing it up
docker compose restart api

# send streaming ticks (from tools container)
docker compose run --rm pytools bash -lc "pip install -q kafka-python && python scripts/produce_ticks.py"
# in another terminal, hit the API while ingest updates Redis
curl -s -X POST http://localhost:8080/v1/score -H 'Content-Type: application/json'   -d '{"symbol":"BTC","ts_ns":1,"features":[0.1,0.2,0.0,0.3,0.1,0.0,0.2,0.1],"freshness_ms":10}'

# debug helpers
curl -s "http://localhost:8080/debug/triton?n=8" | jq
curl -s "http://localhost:8080/debug/features?symbol=BTC" | jq

Notes

• Redis key format for streaming is simple: te:spread_bps:{SYMBOL}. TTL is SPREAD_TTL_SEC (default 180s).
• Redpanda auto-creates the ticks topic on first publish.
• Grafana dashboard (v1.3) will also chart canary metrics if you add them.