MEASURED WORK / PUBLIC RESEARCH
Technical work you can inspect.
These are lab and hardware studies—not invented customer stories. Published studies include the hardware, measurement path, constraints, and public evidence needed to understand what the numbers mean.
MEASURED · PUBLISHED
Making Large Models Fit Smaller Hardware
A 27B checkpoint across a 24 GB Turing workstation card, 48 GB Ada, and DGX Spark. Model fit, quantization, serving stack, decode, prefill, and quality measured rather than assumed.
OPTIMIZE · BENCHMARKS →
MEASURED · PUBLISHEDServing a 307 GiB Model on Hardware You Own
Multi-node private inference across DGX Spark and H200: tensor parallelism, runtime fixes, CUDA graph behavior, speculative decoding, batching, and capacity tradeoffs.
PRIVATE INFERENCE · MULTI-NODE →
IN DEVELOPMENT
Private AI for Regulated Environments
A future architecture study covering trust boundaries, controlled data flow, identity, logging, updates, and operations. It will not link to another study until it is actually published.
ARCHITECTURE · REGULATED