In early October 2026 I posted a series of LinkedIn updates describing practical AI‑engineering tactics such as cheap confidence‑based routing, eval‑driven development, parallel fan‑out pipelines, GPU fundamentals, SVG diagram generation, and the DevPulse capstone agent.
Using Jev for pass/fail scoring can make those decisions 10-100x cheaper than using an LLM.
I used asyncio.gather() with a Semaphore(5) so 5 documents could run concurrently.
Decode uses a semaphore to cap concurrency at 4 subagents.
The DevPulse project reached 85 % test coverage and the SVG diagram skill improved from about 80 % quality to 100 % after several weeks of iteration.