Jason Normore

cevolve. Genetic algorithms for autonomous code optimization. The LLM imagines ideas, evolution discovers which combinations work best together.

unitask. Disposable unikernel powered MCP and CLI tool for untrusted code, with declarative policy.

cognit. A local LLM that doesn't pay for context the way frontier models do. Hybrid Mamba+attention base: fixed-size session state, MB-scale persistence, minimal prompts. MacBook-friendly, OpenAI-compatible.

csa. Calibrated Sparse Attention: a training-free inference-time procedure that reduces attention compute in pretrained transformers by an order of magnitude while preserving language-modeling quality.

pi-evolve. pi extension for evolutionary code optimization using genetic algorithms. Companion to cevolve.

emerge. Ensemble inference proxy for LLMs. Combine multiple small models to match frontier-model accuracy at a fraction of the cost and latency

cue. The runtime where agent-authored apps live. Push over MCP; call by cron, webhook, or another agent — each call runs in a fresh sandboxed VM.