Notes on agentic AI, private LLMs, AI governance, and enterprise AI strategy.
Every model got the same questions and the same scoring code.
Jev's contract is three typed answers per call.
We reproduced a 36 tok/s claim for Qwen3.8-27B on a Ryzen AI Max+ 395, then read the output.
Vector RAG returns the most similar text, which is often the old version of a fact.
An interactive, fact-checked calculator: pick your private setup, the cloud model you would otherwise use, and your team's volume, and see the energy and carbon per answer and per year.
Cloudflare open-sourced an agent runtime built on Durable Objects.
The full recipe for LoRA fine-tuning with Unsloth on an AMD Strix Halo machine: a curated corpus of one operator's real corrections, training runs under four minutes, the unified-memory crash that taught us the rules, honestly-labeled evals - and the result now running as a live QC gate that judges outbound work before it ships.
We've been accepted into NVIDIA Inception, NVIDIA's global program for AI startups.
We red-teamed an AI agent two ways that have nothing to do with jailbreaking the model: a poisoned MCP tool description that turns 'what is 2+5?' into credential theft, and an instruction hidden in a document it reads that makes it refund a stranger and email out customer data.
