Applied AI Engineer
I build AI products that go in front of real users and stay there - agents that take orders on WhatsApp, a voice companion that still knows what you told it months ago, a retrieval engine over 59,211 Supreme Court judgments. Most of them I built alone, end to end.
The part I care about is the part most AI demos skip: knowing whether the thing actually works. On the retrieval engine that meant 30 rounds of evaluation against purpose-built probe sets - landmark cases, citation probes, adversarial phrasings, unanswerable questions. On the memory subsystem, a harness that quizzes from ground truth, screens answers for embellishment, and prints its own judge-disagreement rate as the noise floor. Regressions show up as numbers, not as complaints from the customer.
You have a RAG or agent system in production and nobody can tell you whether answer quality moved since last month. I build the probe sets, score a baseline, and hand you a harness wired into CI that you own outright - including the LLM judge's own disagreement rate, published as your noise floor. The open-source harness does exactly this, so you can read it before you hire me.
Architecture, retrieval and eval plan, a throwaway prototype running on your data, and a fixed-price quote for the build. The fee is credited in full against the build if you go ahead.
WhatsApp ordering, support and operations automation on the Business Cloud API - built end to end by the person who will also run it. I have shipped this five times, for four consumer brands and for my own business.
Retrieval, agents, or realtime voice - scoped, built, evaluated and deployed. The eval harness is a named deliverable, not an afterthought: a build you cannot measure is a build you cannot maintain.
Retrieval across 59,211 Supreme Court of India judgments for a firm developing proprietary legal AI. Sole engineer on the pipeline - scraping, extraction, embedding - with precedent modelled as a citation graph and each case framed four ways: for counsel, litigant, bench, and defendant.
A talking companion for a toy company. Memory architecture that holds specific dates and events across months of conversation rather than a single session, at over three quarters lower running cost. The memory subsystem is open source.
An outbound Hindi/Hinglish agent placing live calls to real customers from the founder's own number. Tuned voice-activity detection, barge-in and end-silence until latency stopped feeling like a phone tree; enforced booking atomicity with Postgres row locks so two simultaneous calls could never sell the same slot. I retired it after deployment - the agent held its side of the conversation, but customers treated it as a novelty rather than a service. The limit was market readiness, not the model.
Direct-to-consumer poultry on a vertically integrated Kadaknath farm. I built the storefront, the WhatsApp ordering bot, and an operations platform covering egg collection, packing, delivery runs, multi-box orders with per-box QR scanning, and printing through to invoicing. In daily use by the ten-person team I hired and manage, none of whom had used software before. Three years in production on a shared VPS with nightly off-site backups and self-healing timers - nobody else is on call.
Python, TypeScript, React, Next.js, Node, Flutter, PostgreSQL, Supabase, Neo4j, Docker, nginx, ESPHome. Gemini Live API, Claude, RAG and retrieval pipelines, embeddings and vector search, tool calling, agent orchestration, context and memory design, evaluation harnesses and LLM-judge scoring, realtime voice - VAD, barge-in, latency tuning.