TiGrIS, a tiling compiler that fits ML models onto embedded devices
Fits 4.5MB models into 256KB SRAM via tiling with zero dynamic memory allocation.
One pixel. Three weights. Real inference. AI model that fits in a single pixel.
Stores a trained model in a single pixel, but only handles binary classification tasks.
ML educators, hobbyists
Fits 4.5MB models into 256KB SRAM via tiling with zero dynamic memory allocation.
Outperforms existing open-source injection detectors on ProtectAI and Qualifire benchmarks.
Custom noise node implementations finally close the gap between reference and open-source renders.
Komilion turns model sprawl into a cost-control layer you drop in by swapping a base_url: requests are classified (regex fast path + tiny LLM) and matched to ~400 models so cheap models handle the easy stuff and premium models only run when needed. The ~60% zero‑call regex fast path and benchmark-driven routing (LMArena) are clever, pragmatic moves; the hard questions left are model-quality drift across providers and how routing decisions map to real-world user satisfaction.
Overlay diff mode shows exactly where each AI model diverged from your design.
Cross-architecture knowledge transfer via activation tokens, but benchmarks lack rigor and baselines.