Writeups from the people doing the work — air-gapped deployment, tamper-evident audit, and the unglamorous engineering that turns a demo into a system the mission can depend on. No hype. The things we would actually send a peer.
We publish when we have something true to say — one topic, drawn from running systems, written for the engineer who has to make it work.
A demo that runs once in a Jupyter notebook is not a system. The gap between the two — eval suites, version control, observability, rollback, human-in-the-loop, audit — is where most federal AI programs stall.
Read →A working pattern for agents where the network ends: vendored harness, local NER for PII redaction, self-hosted inference on Ollama or vLLM, and a run-and-dump container that makes zero external calls.
Read →Which version of which policy made which decision, on what inputs, and what did a human do about it. Tamper-evident provenance is not compliance overhead — it is what lets a government system run, and reach an ATO.
Read →If your AI capability is stuck short of production, or has to run where the network ends, that is the work we do. Operators, not advisors.