An Implementation Field Guide to Measuring AI Work Beyond TokensDisclosure: This is an AI-assisted, human-reviewed field guide prepared for Velokey. The linked first-party resource hub is the Velokey blog; the code and schema remain provider-neutral and reproducibJul 23 · 11 min readvelokey.hashnode.dev
How to Evaluate Gemini 3.6 Flash Without Letting Benchmarks Pick Your DefaultA model launch lands in the engineering chat. Within minutes, the benchmark table is being treated like a migration plan. Gemini 3.6 Flash makes that temptation unusually strong. Google’s July 21 releJul 22 · 11 min readvelokey.hashnode.dev
A Trajectory-Aware Control Plane for Long-Running AI AgentsA long-running agent hits a denied network request. The action guard worked, so the system records a successful block and lets the session continue. That sounds reasonable until the agent treats the dJul 21 · 10 min readvelokey.hashnode.dev
The Capacity Gate I Would Add Before Launching an AI Agent ProductWhen I review an agent launch plan, I picture a very ordinary Friday morning: a product demo lands, sign-ups jump, and every new user starts a workflow that can call a model, invoke tools, retry, and Jul 20 · 8 min readvelokey.hashnode.dev
Build a Five-Gate Evaluation Harness for Kimi K3 and Claude Fable 5A leaderboard moves on Thursday. By Friday, someone asks whether the default model should change. The fastest safe answer is not another opinion. It is a small evaluation harness with explicit workloaJul 18 · 8 min readvelokey.hashnode.dev
