Build and run ATLAS.
Start with the runnable ride flow. Use the linked files, API contracts and acceptance checks to extend it with a coding agent.
01 / START
Run the full demo stack
Use Docker Compose for PostgreSQL and Kafka. The model setup trains on CPU and stores versioned artifacts before serving starts.
git clone --branch architecture-explorer --single-branch https://github.com/FUllSCIPTS/ATLAS.git
cd ATLAS/architecture-explorer/starter
docker compose up -d --build
docker compose exec -T gateway python -m tests.smoke
docker compose exec -T -e KAFKA_BOOTSTRAP=kafka:9092 gateway python -m tests.replay
Open Passenger app and Driver app after startup. Book, accept, pick up and complete a ride.
The starter also includes a Python-only SQLite quickstart. Local run instructions
02 / TRAIN AND SERVE
Run the ML pipeline
- Demand LSTM
- Train on ordered counts and predict the next five-minute count per zone.
- Acceptance model
- Return a logistic regression probability for each driver from distance and speed.
- Anomaly model
- Use Isolation Forest to flag unusual telemetry and log it with the offer.
- RL policy
- Learn from completed rides, evaluate a versioned candidate, then explicitly approve or roll it back.
03 / VERIFY
Know when it works
- A retried booking key returns the same trip and emits one booking event.
- Only the offered driver can accept. The trip reaches completed in order.
- The recorded offer shows a model version, forecast, acceptance score and anomaly flag.
- Completion creates one experience record. The RL approval gate and rollback work.
The starter uses demo accounts and synthetic training data. Real riders require production identity, GPS and routing, re-matching, payments, and evaluation on representative data.
04 / HAND OFF
Give this brief to a coding agent
05 / ANDROID APP
ATLAS Go 1.2.0 on the phone
The Android app is separate from this starter. It has passenger, driver and food-seller accounts and a live map, stores everything on the phone, and ships what the agents learned as plain Python constants.
- Dispatch policy v1.4
- REINFORCE softmax over ETA, acceptance and surge features; registered drivers online on the phone compete with the simulated fleet, and each booking stores a plain-language reason.
- Acceptance model
- A Beta-Bernoulli estimate per registered driver, learned from their own accept/decline history, feeds the dispatch policy.
- Courier timing
- A Q-learning table decides how long to wait before sending a courier, so food and courier arrive together.
- Surge outlook
- Hourly demand ratios learned from two simulated years of Cork over Ireland's public holidays, with the hours the supply agent adds drivers.
- Learning lab
- Retrains a policy from zero on the device and compares it with the nearest-driver rule. It never dispatches.
cd android-app
python -m pytest tests # 117 tests
python training/export_learned.py # regenerate atlas_go/learned.py (needs numpy)
buildozer android debug
Explore the app architecture · ATLAS Go source ↗ · Download 1.2.0