01 — About
Grounded in shipping, curious about breaking things.
I'm an AI engineer who owns systems end-to-end — from architecture and model selection through deployment, monitoring, and the on-call reality of keeping an LLM system correct under real constraints: latency, cost, and trust. My work spans retrieval-augmented generation, LLM inference optimization, and multi-cloud model orchestration across AWS Bedrock, GCP Vertex AI, and Azure OpenAI, shipped and operated in production, not left in a notebook.
Outside of building, I run independent security research on LLM applications — probing production chat systems for prompt-injection and policy-bypass vulnerabilities, and reporting findings through coordinated disclosure. I think the two disciplines sharpen each other: you design better guardrails, and take deployment more seriously, once you've tried hard to break what you shipped.
What I'd bring to your team
- I ship LLM features end-to-end — architecture, model choice, evaluation, and deployment — so a team spends less time handing work across roles and more time shipping.
- I design for correctness under real constraints, not demo constraints: citation grounding, temporal validity, and hallucination rate as release-blocking metrics, not afterthoughts — see the numbers in the NBFC Compliance Intelligence case study.
- I bring an adversarial mindset to what I build, from independently finding and responsibly disclosing an LLM safety vulnerability — the same instinct that catches failure modes before a customer does.
- I default to measuring, not asserting: every claim on this site links to the evidence behind it.
02 — Skills
Toolbox
03 — Featured Work
Selected Projects
NBFC Compliance Intelligence
Retrospective RBI-compliance auditor for Indian NBFC lenders, deployed and actively iterating
An AI-assisted compliance engine that ingests a lender's own loan documents and collections-call transcripts, extracts structured facts, resolves which RBI regulations were legally in force on the date of the event, and issues a cited, auditable verdict — compliant, violation, ambiguous, or no-clause-found — with every citation validated against a versioned regulatory corpus before it's ever shown to a user.
Four independently-testable stages (extract → retrieve → decide → validate), zero raw-document persistence by architectural constraint, row-level tenant isolation, and a production incident I diagnosed and fixed myself: sequential per-fact LLM calls were blowing past reasonable timeouts, fixed by fanning them out concurrently for a ~4x wall-clock improvement, with a regression test so it can't silently regress.
AI Safety Vulnerability Disclosure
Responsible disclosure to a major LLM chat platform
Identified and responsibly disclosed a prompt-injection vulnerability affecting safety-policy enforcement in a major LLM chat product. Submitted a coordinated disclosure report to the vendor's security team, including reproducible test cases, supporting evidence, a technical writeup, severity assessment, and mitigation recommendations.
04 — Contact
Let's talk.
Open to Applied AI Engineer / LLM systems roles. Reach out directly — I reply fast.