I build products, and then I build the layer the agents running them turned out to need. The order is the whole method. You can reason about what an agent needs to operate software, but you only get the real list by running something and finding what it can’t reach. What I learn doing that I keep as eight doctrines at Tinkerers Labs, the name I ship my own projects under.
Founding Engineer at Kay.ai, where I own AI products that do insurance work in production. Some of that is policy checking: a customer sends last year’s policy, its endorsements, the ACORD application and the renewal, and the work is finding small inconsistencies across four documents that are all supposed to agree. It is the least forgiving place I could be doing this, which is the reason it is useful.
Profilebud, founding team. I led frontend and wrote a Go server handling 20M users’ data at 1K req/s. It reached $7K MRR before Meta shut it down with a cease and desist in July 2021.
EngageBud, co-founder and CTO. Built Influencerbit and Engagebud, passed 13M users, raised $100K from Upekkha.
Dreamboat.ai, co-founder and CTO. An LLMOps platform, with a Cloudflare Workers proxy holding sub-20ms.
Since then, things I still run: fetchbean, niahere, aijobsdesk, Promptsmint, and the template this site runs on. Full timeline · Everything shipped
Mostly the explanation I wanted and couldn’t find. Embeddings in four parts. What quantization costs you. Why one model name behaves differently across providers. Agent skills from scratch. Some of it isn’t technical: ten days from Leh to Srinagar, a year in review. All of it here.
Bengaluru. Techno and house. F1, chess, and reading. Last trip was Ladakh, before that Bali, and a month across Thailand and Vietnam.
Easiest thing to email me about: something you’re building that an agent has to operate.