I build AI-driven products end to end: React, Next.js, and React Native on the frontend, Node.js and NestJS on the backend, and multi-agent AI systems tying it together. I have professional experience building and shipping production software, from cross-platform mobile apps to secure backend services, plus three AI products built and shipped on my own, with more in progress.
The differentiator: designing and directing AI systems, not just calling an API.
An AI-native fitness and nutrition tracker: a conversational agent backed by a single SQLite database, with no server and no separate app. Built as a 16-skill modular agent system, with cost-driven model routing bringing projected running costs down to $1–2/month.
An AI-assisted quiz learning platform with a React frontend and dark/light themes, and a NestJS backend generating dynamic quiz content through the OpenAI API. A caching layer detects repeat inputs and reuses quizzes instead of new AI calls, cutting API costs early on with much more headroom as usage grows.
A multi-skill AI system that generates job-tailored CVs as compiled PDF documents. It reads a job posting, extracts role focus and must-have keywords, then reorders and re-emphasizes only your existing experience and skills to match, with nothing fabricated. Fully automated from posting to a ready-to-send PDF.
A 3-step Claude Code plugin (init-ad, create-scenario, create-ad) that generates production-ready animated HTML5 ads directly from an app's branding and feature set. Brand style and features are configured once and reused across future ads with no re-entry.
Can a language model learn chess just by reading moves? A 5.8M-parameter GPT-style Transformer I built from scratch in PyTorch and trained on 289,888 games, where every move is one token in a 1,932-token vocabulary. It never sees a board and gets no help from a chess engine, yet it picks a legal move 91.6% of the time. Linear probes on its hidden states show it builds a picture of the board inside its 6 layers.
How can Real Madrid stop Barcelona's dominant La Liga run? I used an AI agent to gather and cross-check public match data on Barcelona's 7-0 start to the 2026/27 season, compared it with 10 dominant win streaks that were broken between 2015 and 2026, and built a game plan for Real Madrid. Presented as a narrated, animated video built from scratch in HTML, SVG and JavaScript.
Is a League of Legends game really lost at 15 minutes? A data analysis project testing a common belief in competitive gaming. I built a dataset of 100 ranked games from a Challenger-level player, analyzed it with SQL, and presented the findings as an interactive one-page React site.