Took an applied-AI platform for residential real-estate investment from concept to production as the sole engineer — React frontend, multi-agent backend, Azure infrastructure — later directing two contract engineers.
Cut public-health reporting latency for state health departments from 24 hours to near real-time, migrating the National Electronic Disease Surveillance System off batch onto an event-driven Kafka pipeline.
Replaced a ticket-driven permissions process with a self-service portal, taking customer wait time from seven days to three hours, and moved a Tier 1 service off legacy DNS onto Route 53.
Automated PTO request and approval workflows in an internal CRM, cutting over 1,100 hours of manual processing a year.
Everyone’s instinct is to hand the whole matching problem to the model. We gave it the messy half and kept the ranking deterministic — here’s the argument, and the part we still haven’t solved.
An agent in production failed in a way no eval suite would have caught — because the thing that changed wasn’t the model, and the thing that broke never threw an error.
Heads down on Rehouzd. v0 buyer-matching in the hands of first wholesalers.
- building
- tooling that turns messy off-market real-estate workflows into clean operator surfaces
- reading
- Atomic Habits · The Hard Thing About Hard Things
If you’re hiring, building something adjacent, or just want to argue about agent reliability — the inbox is open.
ragulshanmugam@yahoo.com