Hardware-first observability and cost governance for self-hosted inference. A WES efficiency score (tokens per watt) and real cost per million tokens across your own GPUs. Free community tier.
A test harness that checks whether data in your AI stack actually stays inside the border you drew. It runs a canary through the real stack, stresses it to force failure and fallback behavior, then inspects every layer data lands in (logs, caches, vector stores, egress) and hands you audit-grade evidence. Residency proven by behavior, not by diagram. The first tool in a proof layer for sovereign AI.
An open methodology (not a commercial product) for training small private
enterprise models. Current work: account-intelligence-7b-v1, a fine-tune of
Qwen2.5-7B for enterprise account briefs, built to run on owned hardware and
emit structured signals rather than prose, as a component in compound-AI
pipelines.
An internal agentic enterprise platform running Qwen2.5-32B-FP8 on a DGX Spark alongside Claude, generating account intelligence, meeting prep, and pricing assistance for sales teams via integrations with Salesforce, Confluence, and other internal systems.