Currently working on: routing logic/pipelines and fine-tuning small encoders trained on that logic.
Hive — a local-first stack for memory, verification, and routing in language models. The repo holds Splinter Memory, The Hive Cascade, and HiveBench, the evaluation suite that measures both. Papers below · Visual companion.
Scion — source, notes, tests, parameters, calculations and formulas behind the Scion series of LLMs I quantized and fine-tuned on rented GPUs and locally, all recorded. Scion-35B-A3B and Scion-FlashNext-176B-A6B.
PixelQR — pixel art that becomes a real, scannable QR code. Live on the site.
OpenFG — MIT-licensed optical-flow frame generation for Linux: an original C++/GLSL backend and Vulkan layer with 2×/3×/4× insertion, same-GPU and cross-GPU DMA-BUF, and adaptive target FPS that won't start FG until below the threshold.
- Splinter Memory — managed-decay context curation for long-horizon conversations: the bounded context, the encoder fleet, the measured verdicts and the negative results.
- The Hive Cascade — a gold-free verification-and-repair loop for local models: solve → detect → repair → escalate, with the controller contract and the measured tier table. Visual companion.
LLM & ML
Systems & performance
Web & tooling
Data & testing
Hive is public. Some engine and experimental work remains in private repositories while it's in progress.