Kenoma Labs builds autonomous decision systems for environments where the state space is enormous, feedback is delayed, and mistakes compound. Markets, adversarial games, resource allocation under constraint.
Acting well in complex environments looks less like prediction and more like search. Every decision opens a tree of futures, and most branches are bad. The real work is pruning fast enough to matter.
We build agents that learn policies through interaction with high-fidelity simulations, decompose hard problems into tractable subgoals, and use the structure of their own failures to improve. Less "predict the answer," more "learn to act when you can't."
Kenoma Labs is self-funded by design. No investor timelines, no pivot pressure. We pick problems for their difficulty, not their market timing, and stay on them until the agent works.
Reinforcement learning systems trained in adversarial, partially observable environments — from limit order book dynamics to multi-agent market microstructure — where latency and sequencing define the edge.
GPU-accelerated simulation engines that replay and extend real-world sequential data at scale. C++ core with Python orchestration, designed for millions of decision points per training session.
Deep learning systems optimized for throughput at the hardware level — distributed training pipelines, kernel-level inference optimization, and model compression for latency-critical deployment.
Architectures where agents generate their own tooling, evaluate their own performance against ground truth, and systematically close capability gaps without human-in-the-loop intervention.
The kenoma is the void — the space of deficiency, where what should exist does not yet.
This is the landscape of every unsolved problem. Unexplored strategies live in the kenoma. A working policy is a small act of filling some of it in.
We named ourselves after the territory.