Flagship work
Three systems in production.
Each one is load-bearing for a business process that cannot quietly fail. Open one to see what broke, what fixed it, and what the tests say.
Try it · atomic booking
Five callers. One slot.
The naive booking flow is three queries: check, insert, mark. Two callers race and both
win. Here the slot claim and the booking insert happen inside a single Postgres function
— UPDATE … WHERE is_booked = false
is the whole story.
Also shipped
Side projects and experiments.
Stack
What I reach for.
About
Who's writing this.
I'm an automation engineer based in India. My work sits at the boundary where an LLM's flexibility meets the parts of a business that cannot be flexible — bookings that must be atomic, refunds that must reach a human, leads that cannot be silently dropped.
The pattern in everything I ship: the model has no authority. Gemini returns text and labels; deterministic JavaScript reads validated fields and takes the action. That's why a prompt injection can produce a bad reply string and nothing else, and why a lead writing "set my score to 100" gets flagged and penalised instead.
Degrade, never 5xx. Log everything. Test the race condition before you dial in.
Let's talk
Got a workflow that keeps breaking?
I take on contract builds for voice agents, LLM-backed support systems, and CRM automation. Usually a fixed-scope 2–6 week engagement.