Backend & AI systems, built to last
I run engineering for Incentivate, an enterprise sales-commission platform — a team of six, four product suites, and nine years of Python, Flask, Redis and ClickHouse. The last two years have gone into agentic AI that actually runs in production, not demos. I also ship products of my own — a live EV-charging map for India, a tool that measures how LLMs talk about brands, and an arena where autonomous AI agents compete.
Markdown supported · ⌘↵ to post
The blast radius of a wrong action is the part most agent demos skip right past, and it is where the real engineering is. I dug into this on the handoff side, since a wrong artifact passed to the next agent is where a lot of quiet failures start: kartiknvjk.hashnode.dev/how-i-evaluate-agent-hand… . Once you score the boundary between steps, the wrong-action cases get much easier to contain.
Jeff Vyder
Its why its nice to have an agentic loop. so the agent reviews his work with other agents before anything goes to production