Hi, I’m Vishal Patil.
I build AI products people actually use.
AI engineer and researcher shipping end-to-end products from specification through production. My last project, VAi - an AI assistant that talks to recruiters on a job seeker’s behalf - went viral and was featured by CNBC.
I care about the whole loop: talking to users, freezing a spec, choosing the right stack, wiring the eval harness, and shipping something that stays up. My philosophy is baked into Throughspec - a spec-driven SDLC for Claude Code that refuses to skip cross-questioning, keeps memory files under 8k tokens, and produces reproducible architecture decisions instead of chat-log folklore.
On the applied side, I have built retrieval pipelines over messy real-world corpora, evaluation harnesses that catch regressions before they ship, and multi-provider LLM routing that keeps latency and cost predictable. I care about security defaults - GDPR-by-design, PII minimization, audit trails - and I write code the same way I write specs: boring, deterministic, tested.
If you are building an AI product and want someone who can own it from cross-questioning to production, let’s talk. My best work happens when the spec is honest, the memory layer is condensed, and the structure carries its own weight.
Stack I ship with.
Ship features with SRS-first workflows; the practice Throughspec itself is built on.
App Router, server components, static export, edge-friendly builds.
Retrieval pipelines, vector stores, embedding evaluation, hybrid search.
Fallback routing across Anthropic / OpenAI / open-weights; cost + latency tuning.
PII minimization, request-scoped scoping, audit logging, DPIA-ready defaults.
Schema design, migrations, row-level security, connection pooling under load.
Shipped work.
Let’s ship something worth remembering.
Full-time, contract, or advisory. Open to work across the US & Europe.