Custom engine training

We build neural-network AI for board games — from network architecture to production inference.

We built our own engine for backgammon and Long Narde: networks train via self-play, get vetted in tournaments against prior generations, and only promote to production when stronger. The same pipeline transfers to other games.

Our key edge is splitting game phases (blocking / race / asymmetric) into dedicated networks trained on stratified positions. That delivers accuracy where one-size-fits-all networks lose ground.

Tech stack

TD-Gammon style

TD(λ) training, no hand-labeled data. Millions of self-play episodes via a GPU rollout server.

Multi-phase nets

A dedicated network per phase (blocking, race, asymmetric). Automatic phase routing per position.

ONNX inference

Models exported to ONNX. Inference via the ort crate on CPU / CUDA / OpenVINO. Tens of ms per position.

Distributed training

Rust pipeline: rollout → train → tournament → promote. Supervisor daemons, checkpoints stored in the DB.

Equity & analytics

Ready-made infra for PR rating, equity loss, move classification. Plugs into the engine one-to-one.

Production-ready

End-to-end from self-play to production inference. Running for a year in our own consumer app.

How we work

  1. Discovery: rules, game phases, evaluation specifics.
  2. Network architecture: inputs, phases, outputs, sizes — iterated on small runs.
  3. Self-play training: run a generation chain on GPU, monitor progress via tournaments.
  4. Promote and integrate: final models → ONNX → your inference stack.
  5. Support: new generations, tuning, edge-case improvements over time.

Discuss your project

Tell us about the game, target timeline and intended use (consumer app, research, embedded). We'll get back within a couple of business days.