TALASON
AI TRAINING GYM

A place for agents to learn.

Configurable simulation environments for training and evaluating AI agents: scenarios, sensors and physics you control, with a clean interface to your training stack and full runs on local hardware.

What we build

The pieces we deliver.

Environments

Scenario definitions, procedural variation, and physics from a custom C++ engine or Unreal, exposed through a simple step / observe / reward loop.

Sensors and observations

Cameras, depth, ray casts and state vectors, recorded in sync so datasets are reproducible.

Local inference

Multi-GPU orchestration for local models with llama.cpp and Ollama, benchmarked and logged, no data leaving the building.

Evaluation

Replay, telemetry and metric logging so you can see what the agent actually did.

How we work

Plain process, working builds.

SCOPE

A short call and a written plan: what gets built, in what order, and what a first playable or runnable milestone looks like.

BUILD

Weekly builds and written progress reports. Source in your repository from day one.

HAND-OVER

Documentation, profiling notes and a team that can keep going without us, or with us on retainer.

Related work

Where this comes from.

Local LLM orchestration toolkit

Shell and Python tooling for running, benchmarking and managing large local models across mixed GPUs.

Internal

Custom C++ simulation engine

PBR rendering, dual physics backends, ECS and an integrated editor, built as the base for training environments.

Internal

Have a project in this direction?

Send a short note. We reply within a couple of days with questions or a first take on scope.

Start a conversation