I'm 18, and I'm building Flying Police — AI security for the physical world, with tamper-proof, on-chain incident logs. Along the way: 3× hackathon winner, three startup internships, and twenty-plus products shipped this year.
The short version
I'm 18. In the last year I've shipped 20+ products, won 3 hackathons, and started Flying Police — AI security for the physical world — mostly solo. I don't wait for permission to build. I build, break it, and ship the fix.
Currently building — Flying Police: on-chain, tamper-proof incident logging so a security feed can't be quietly edited after the fact.
Building agentic workflow systems — autonomous multi-step execution rather than assistance — at a holding company that spins up independent AI companies around markets where AI can own an entire workflow end to end.
Built a real-time AI voice-translation tool (speech-to-text → translation → text-to-speech) as a standalone product, taking it to a near-complete working prototype. Also improved SEO across web content to grow organic reach.
Supported the research team in training and evaluating a healthcare AI model, contributing on the data and experimentation side — concurrent with the FlytBase internship.
AI security for the physical world — turning any camera into a guard that never blinks. It analyses footage with a vision-language model, fires deterministic alerts, and writes tamper-proof, on-chain incident logs on Monad. Built for factories, warehouses & industrial sites.
A personal agent for college and the C3A Labs internship, living on Telegram — with Jake Peralta's personality and non-negotiable accuracy. It scrapes Moodle for deadlines and grades, mirrors the institute's own attendance figure, and does the recovery maths on what skipping a lecture actually costs. Hourly polling, escalating alerts, a 07:30 brief, and a confirm-before-acting gate so it never acts unasked.
The origin of Flying Police — AI remote-security monitoring for factories & warehouses: frame-by-frame vision-language analysis, deterministic alerts, real-time Telegram notifications, and a Q&A interface over the footage. Won LocalDev's Codex Community Hackathon, Pune.
The open data layer for Physical AI. A browser game where steering a robot arm silently generates real, LeRobot-compatible robot-training data — then an in-browser policy replays what you taught it.
An agentic system reuniting families separated at the Kumbh Mela (80M+ pilgrims): multilingual intake → semantic matching → safety-verified coordination. Exposed as an MCP server, with Polygon-anchored records.
A live arena where LLM agents play a strategy game — trade, ally, attack, betray — with their reasoning streamed live and a judge LLM calling winners. Provider-agnostic across Groq, Gemini & Claude.
An autonomous pipeline that scrapes Devpost & Unstop, deduplicates listings, and ranks hackathons by personal fit using Claude — delivered as a cron-scheduled email digest.
A voice companion that lives in the Mac menu bar. Clap twice, say her name, or hold ⌥ Space — she answers out loud in about a second and a half. Built around a persistent four-dimensional mood (valence, arousal, warmth, concern) that decays toward baseline and is nudged by what actually happens on the machine.
A second brain rendered as a holographic web-orb you move with your hands. It reads E.V.'s accumulated memory — facts, episodes, journal, mood — and lays it out as one shape, while Claude Opus 5 answers over it and drives the visualisation as it speaks, lighting up whatever it's describing.
A knowledge vault generated from what is actually on the machine — real projects and running agent systems, never demo data. Full-text search, link/staleness diagnostics, a 3D force-graph view, and a JARVIS voice layer over the top. Every writer checks a disowned-facts list before it can write.
🛟 Multimodal drowning detection: Gemini vision on the pool feed, a predictive risk engine that escalates before an incident, an autonomous agentic Incident Commander, and Telegram alerts to the humans who can act.
Know your chances before CAP does. An independent reconstruction of COEP Technological University's B.Planning admissions, built entirely from the official published record — every allotment list, vacancy notice and cut-off table parsed and cross-checked into a prediction engine. Where the record can't answer, the app says so.
A from-scratch, loop-native personal agent — not a chatbot you open, but a loop on a heartbeat that finds what matters, drafts the move, has a separate agent check its work, remembers everything, and surfaces only what's worth your attention. Five primitives, hand-built, no agent frameworks.
A privacy-first, local-first cockpit for taking back your digital footprint: surface the accounts exposing your data, draft legally-grounded erasure requests in seconds, and track every removal until it's actually done. No scraping, consent-based only — tracker data never leaves the browser.
help, ls, cat resume, open jake — arrows for history, Tab to complete.