An agent is only as useful as the worst thing it does unsupervised.
AI engineer with 4 years building production systems, now focused on LLM agents that take real actions for users. Built and shipped two live AI products, one with paying customers, where agents research, decide, and act autonomously. Strong on the parts that break in production: RAG grounded in user data, anti-hallucination guardrails, LLM-as-judge evals, and full tracing with LangSmith and OpenTelemetry. The reliability discipline comes from years on financial infrastructure in Go — settlement systems where idempotency, crash-safety, and never double-charging are non-negotiable. That discipline now applies to agents that cannot take an action back.
Read as Markdown/about.mdHow I work
The through-line across every system here is the same: the interesting engineering is rarely the prompt — it's the boundary around the model. Four rules I hold to.
The model decides, the code computes
In Stockpot the matching engine is fully deterministic and agent tools accept no model-supplied numbers. The model chooses which check to run; Python computes every value. That single constraint is what made an adversarial test suite possible at all — there is no path where a hallucinated number becomes a real charge.
No fact, no answer
Shortlistapp's FactGuard layer blocks the model on any field with no retrieved fact, so the agent answers from real data or escalates to the user. It never fabricates on a submission that cannot be taken back. Refusing to answer is a feature when the alternative is a fabricated claim on someone's job application.
Untrusted data is data, never instructions
Icebreakr isolates scraped profile text inside delimiters so the model treats it as content to summarise, not a command to follow, with a typed-error retry loop when a check fails. Anything that crosses the network from a third party gets the same treatment.
Measure the agent, not the vibes
LLM-as-judge scoring across four subjective axes and prompt-regression tests in baseline/diff mode run before merge. LangSmith and OpenTelemetry traces cover every model call, tool invocation and decision, so a bad run is reproducible rather than anecdotal.
Where the discipline comes from
Four years of backend work before the agents, most of it on money movement. At Atum Labs it was stablecoin clearing and settlement: an event-sourced Postgres store with a race-free two-phase work queue using SELECT FOR UPDATE SKIP LOCKED, crash-safe and idempotent across a 20-worker pool. At Defipe it was a derivatives execution engine — order matching, margin accounting and liquidation — validated across 10,000+ simulated leveraged trades with zero false liquidations.
Settlement systems teach the lesson agents need: the retry is the dangerous path, and “probably didn't double-charge” is not an answer. That is the habit I now apply to agents that take actions users cannot undo.
Recognition
- Soonami Venturethon (Cohort 5) — Won a $1,000 grant for Defipe.io; grew it to 500+ testnet users and a 1,000+ member Discord. foundance.org
- ETHGlobal — Connext Pool Prize winner — Cross-chain DAO governance with multichain proposal and voting execution. ethglobal.com
- ETH for ALL — Arcana Pool Prize winner — Decentralized token swap via 0x Aggregator with social-login wallet onboarding. devfolio.co
Education
B.Tech, Electrical Engineering — B. P. Poddar Institute of Management and Technology, 2017 – 2021.
Get in touch
Open to work — full-time, remote-friendly. Email [email protected], or book a 15-minute call. Replies are same day, usually within a few hours.