// the find
kossisoroyce/timber
Ollama for classical ML models. AOT compiler that turns XGBoost, LightGBM, scikit-learn, CatBoost & ONNX models into native C99 inference code. One command to load, one command to serve. 336x faster than Python inference.
Timber compiles classical ML models — XGBoost, LightGBM, sklearn, CatBoost, ONNX tree ensembles — into self-contained C99 inference code, then serves them via an Ollama-compatible HTTP API. The core idea is sound: Treelite has done this for years, and compiling decision trees to straight-line C is a well-understood technique. The claimed 336x speedup over Python XGBoost is plausible for a 50-tree model; latency in the single-digit microseconds for compiled tree inference is realistic.
The zero-dependency C artifact is genuinely useful for edge deployment — shipping a 48 KB .so beats dragging a Python interpreter and a framework onto a gateway device. The Ollama-compatible API surface is a smart adoption angle: teams already running Ollama locally get a familiar endpoint with no SDK changes. Benchmark methodology is at least documented and reproducible with the provided scripts. The model format coverage (XGBoost JSON, LightGBM .txt, sklearn pickle, ONNX tree ensembles) hits the formats people actually use.
The feature list is implausible for a 688-star, 23-fork project: real CUDA codegen, FPGA HLS output, WCET analysis, DO-178C/ISO 26262 certification reports, MISRA-C compliance checking, ROS 2 node generation, PX4 modules, differential privacy, and Ed25519 supply-chain signing do not all ship production-ready in a single pip install from a tiny team. The readme even admits the certification and WCET outputs are 'heuristic / regex-based' — generating fake DO-178C level A reports is actively dangerous if anyone in a regulated industry treats them as real. The entire roadmap is marked done with no meaningful backlog, which is a reliable signal of a project optimized for appearances. Treelite already solves the core problem, has years of production use, and a real community; 23 forks after a launch with this much apparent scope is a bad sign.