About
Engineer,
then adversary.
I'm Madhav Visani, a software engineer based in India. I build LLM applications and the backends that hold them up — then I try to break them, because I only trust a system as far as I have measured it.
What I work on
Retrieval and vector search over a product's own data, structured output the rest of the codebase can actually depend on, failover across providers so one vendor's rate limit is not an outage, and offline evaluation so a prompt change can be shown to have helped rather than assumed to.
Attacks implemented from the papers that introduced them — PAIR, TAP, AutoDAN, Crescendo, Best-of-N — plus indirect injection through RAG, tools, PDFs and images. Findings come out reproducible and classified against the OWASP LLM Top 10 and CWE, not as screenshots.
TypeScript and Next.js on the client, Python with FastAPI or Node on the server, PostgreSQL with row-level security and Redis underneath. Shipped on Vercel, Supabase and Cloudflare Workers, with the auth, the rate limiting and the deploy path included rather than deferred.
Measure, do not assume. Nexus Quant badges every number LIVE, DELAYED, MODELED or SIM so nothing synthetic can be mistaken for real. Market Pulse scores its sentiment model against a human-labelled golden set. Frontier deletes a posting rather than keeping a dead one for traffic.
What I have shipped
Five products, all live, all running in public. Each one has a write-up here and a front door of its own.
- Frontier — An engineering job board and a morning AI briefing. Frontier reads company career pages every thirty minutes and deletes every posting after seven days.frontier.bymadhav.com ↗
- WRAITH — An autonomous red team for AI applications — it breaks your model before someone else does.wraith.bymadhav.com ↗
- Market Pulse — See what your competitors' customers are begging them to fix.pulse.bymadhav.com ↗
- Nexus Quant — A keyboard-first, command-driven market terminal that runs in a browser tab.nexus.bymadhav.com ↗
- Lead Rescue — An AI receptionist that never misses a lead, day or night.leads.bymadhav.com ↗
How I work
- Nothing is trusted that has not been measured. If a system claims a number, something in the codebase has to be able to check it. That is why the evaluation harness ships with the model and the data badges ship with the feed.
- Fail over, do not fall over. Lead Rescue runs five LLM legs across three vendors, so a quota wall is invisible to the visitor, and falls back to a plain form rather than dropping the lead.
- Delete rather than hide. Frontier removes a posting after seven days and returns a real 404 for it, instead of keeping a closed role around because it still earns clicks.
Elsewhere
The code is on GitHub, the history is on LinkedIn, and the resume is one click away. If you would rather just talk, the contact page goes straight to my inbox.