Agents break on software built for people.
I’m Aman, in Bengaluru, and I build the tools that get them through it. Founding Engineer at Kay.ai, where I own AI products that do insurance work in production.
You cannot work out in advance what an agent will need. You get the list by putting one on real work and watching. At Kay the work is insurance, so the list is bad PDFs, legacy carrier portals, and forms nobody has updated since 2009. A person works around all three without noticing they did.
The pieces I end up building usually outgrow the product they came from. fetchbean came out of needing tool access that held up. niahere came out of wanting one agent I could read end to end, rather than a swarm of narrow ones I could not debug. What holds across all of it I keep as eight doctrines at Tinkerers Labs, the name I ship my own projects under.
Shipping since 2020. I co-founded Dreamboat.ai, an LLMOps platform, and EngageBud, which passed 13M users and raised $100K from Upekkha. Before those I was on the founding team at Profilebud, where I wrote a Go server handling 20M users’ data at 1K req/s and got it to $7K MRR, until Meta shut it down in July 2021.
Currently shipping
- fetchbeanOne key in front of 250+ tools, over HTTP or MCP.
- niahereA personal agent you fork and run yourself, small enough to read end to end.
- aijobsdeskAround 35k live roles, read from company hiring systems rather than reposted.
- Tinkerers LabsThe name I ship my own projects under.
- ProjectsEverything I've shipped.
- WritingI write up what I tried and what broke, 72 posts so far.
- ExperienceFull timeline.
If you’re building something an agent has to operate, I’d like to hear about it.
