Johnson James Kanjirathinkal
I'm Johnson, co-founder at Skillfully, a public benefit corporation in San Francisco, where I oversee Product. The day job is multi-agent architecture, LLM evals, and the failure modes that surface when models grade other models. This site is where I write. Mostly about the gap between what we assume things do and what they actually do, in AI and outside it.
Work
Skillfully's simulations run in real time across live voice, adaptive chat, and tool-based tasks, for enterprise customers. Our team built a modular multi-agent framework that keeps long multi-turn simulations reliable.
Before Skillfully, I worked in enterprise machine learning across Asia, on multilingual text analytics and customer-experience platforms, alongside roles at Lawrence Berkeley National Laboratory, Geckolyst and various startups. I hold an MEng in Industrial Engineering & Operations Research from UC Berkeley with a research focus on ML optimization and IP analytics.
UC Berkeley Startup Innovation Award · GSV Cup 50 · Judge, Berkeley StEP, Big Ideas Contest and various others.
Writing
Essays on Substack.
- Nobody demos a blank box
Why open-ended prompts reward users who arrive knowing what to ask. Most people don't.
- The judge gets flattered first
What grading co-authored transcripts revealed about reward signals in LLM judges.
Elsewhere
Contact
I occasionally invest in and advise early-stage AI companies.