Taking systems apart, analysing them, and rebuilding them smaller, better, faster, stronger.
Independent LLM researcher focused on: long-horizon memory, ternary quantization, LLM forensics & analysis, local-hosting on consumer hardware, and more.
Scion — Source, Notes, Tests, Parameters, Calculations and Formulas that helped build a working recipe for the Scion series of LLM's I have quantized and fine-tuned via rented GPU's and local training, all recorded. Scion-35B-A3B and Scion-FlashNext-176B-A6B.
HiveBench — LLM evaluation suite-harness for; A/B testing, re-slottable encoder-sized LLM for harness-based controls like RAG and long-context curation, switchable inference-engine based controls, model fine-tuning, Agent-loop control (API or local).
PixelQR — Pixel art that becomes a real, scannable QR code. Live on the site.
Strata — Kernel fusions and speculative decoding in the Strata inference engine that makes it work with hive-bench (fork work).
LLM & ML
Systems & performance
Web & tooling
Data & testing
Note: Unless otherwise stated, the source code for my current projects is hosted in private repositories for intellectual property and security reasons.