I build AI products that read the person, not just the data.
Right now I'm enjoying the work of turning a messy real problem into an agent that makes good decisions and knows its own limits.
On my own time, with no one assigning it, I pick real problems, design an AI agent for each one end to end, and write up the reasoning behind every choice. It is how I keep learning, and it is the clearest way I know to show how I think instead of just listing skills on a resume.
I like to sit with a problem until I understand what is actually going on, not just the obvious symptom, and then design for the real thing.
Selected work
Shipped product
A journal that writes back, for the ones who always wonder why. You type what is on your mind, Loomavi reads the pattern underneath it, and you get a personal insight in about 30 seconds. Live and in use, free to try.
Design case study
A read-only AI agent that recovers the good customers a blunt fraud rule wrongly blocks. It investigates one blocked-but-low-risk card payment at a time with six read-only tools, then hands the evidence to a two-reviewer board: the same model given opposite jobs, protect the bank versus protect loyal customers, each deciding blind, with plain code reconciling the two votes. It cannot move money, and it escalates anything it is unsure about.
Design case study
A design case study for an AI agent that decides what to do when a grocery item is out of stock and the customer cannot be reached. Instead of a blind refund or a random swap, it reads the customer's intent and picks one of four actions: substitute a close match, source the exact item from a nearby store, hold for confirmation, or refund. It backs off when it is not sure.
Reach me on LinkedIn or at punyaramaswamy@gmail.com.