Not demos. Working systems that run where wrong answers have real consequences, live in regulated use, plus products I have built and run end to end.
End to end. Architecture, prompts, infrastructure, interface, deploy. Solo, or embedded with a small team.
The hard part of production AI was never capability. It was trust. A model that hands you a confident wrong answer is worse than a slow one, because you stop knowing when to check its work. I build systems that flag what they are unsure about and route the rest to a human.
Products designed, built, and run end to end, on my own.
A zero-knowledge password and passkey manager for teams. Save, autofill, and generate logins, store passkeys, and share any login with a team, end-to-end encrypted. Free for 3 people, then $5 a month flat.
A Chrome extension that scrapes business leads straight from Google Maps — name, phone, website, category, reviews — and exports clean CSVs ready for outreach. Free to 40 leads, Pro unlimited.
Phone-triggered remote ignition for Claude Code sessions. Tap from your phone, an agent on the home machine boots a fresh session from zero and hands back a live link. Pull-based, so the cloud never reaches into the house.
Your YouTube, on autopilot. Runsheet plans a channel’s next videos from its own numbers, the creator’s own Claude makes them as motion graphics in one of five styles, Shorts and long videos with chapters, and every title, description and thumbnail is written by YouTube’s rules. It schedules them every week with nobody there, and emails a link to pull any video before it goes live.
A live team Pomodoro product. Shared focus rooms, ambient sound tools, a full marketing site, and three free tools. Designed, built, and maintained solo.
Built for regulated and high-stakes work, six live in production and two voice systems in development. Client names withheld. Same operator, different shapes of problem.
A two-model pipeline reads government RFP PDFs and extracts requirements, deadlines, and rubrics. A second model reviews and flags the extraction.
A rule plus LLM hybrid validates submissions against long guidelines and marks each requirement pass, fail, or needs a human.
Scores cases against a structured catalog of evidence-based therapeutic models, recommends the best fit, and shows its reasoning. The decision stays human-led.
Transcribes and structures calls and documents into case records automatically, inside the existing platform. Removes manual data entry.
Categorizes and summarizes each incoming case and surfaces insights, so caseworkers start with context instead of a raw record.
An ensemble classifier on serverless infrastructure flags AI-generated submissions, surfaced inside the reviewer's case page, with a fallback as detection evolves.
A multilingual voice front-end on a 311 line, all-AWS over a Salesforce routing brain. Routine requests become cases and are deflected; unclear or urgent calls transfer to a live agent — one engine shared with the on-screen console.
An all-AWS generative voice AI on an emergency intake line. Every call is treated as an emergency by default; confident non-emergencies are warm-handed to the right service, and the dispatcher sees read-only unit and ETA data. The AI never allocates a unit.
Tell me what is blocked. I reply to every real message.
kalpesh@buildifyapp.in →