I work with generative AI to understand its product implications firsthand – what works, what breaks, and where the real value is. For my professional background, see About Me.

What’s Next, a movie recommendations app (2026)

A full-stack movie recommendation web app that uses natural language prompt and past data on user’s likes/dislikes to generate AI based recommendations on what to watch next. Built end-to-end using Manus to test how AI development tools reshape the product building process.

How it works: Users sign in via OAuth, rate movies they’ve seen, query for additional movies using a natural prompt and get personalized recommendations. The app integrates with TMDB for movie data, saves user’s ratings for future and uses AI to suggest new movies.

What I learned: The initial version – React frontend, Node.js backend, database, auth – was generated in minutes. But getting the UX right took multiple turns of collecting and synthesizing user feedback. The bottleneck wasn’t building; it was knowing what to build. Wrote about this in The Shifting Bottleneck.

Status: Live at whatsnext.manus.space. Requires Meta account sign-in.

Built with: Manus, React, Node.js, TMDB API, OAuth

Interview Feedback Agent (2026)

An AI agent that helps interviewers write structured, high-quality feedback – reducing time-to-submit and improving consistency across hiring panels.

How it works: Interviewers input raw notes from a session; the agent maps them to Meta’s hiring rubric dimensions and generates structured evaluations.

What I learned: The challenge wasn’t generating structured text – it was calibrating output to match the evaluative rigor hiring panels expect. Prompt design required iteration with actual interviewers, not just test data.

Status: Deployed internally at Meta. Used by multiple interviewers.

Built with: Meta’s internal agent platform, prompt chaining, few-shot classification

Content Agent for X (2025)

An autonomous multi-step agent that generates original product management content – tactical tips, frameworks, and observations on building at scale – and publishes to X with human-in-the-loop approval.

How it works: The agent generates content, routes it for approval, and handles time-based scheduling. It runs as a pipeline with decision logic at each stage, not a single prompt.

What I learned: Autonomous generation works well for structured, domain-specific content. The human-in-the-loop step catches posts that are technically correct but miss nuance – a pattern that generalizes to most AI-assisted workflows.

Built with: Python, Claude API, X API, cron-based scheduling

GitHub

Other Builds

Solo Scrabble (2024): A solo Scrabble game with persistent cloud state and cross-device sync. Start on your phone, finish on your laptop. GitHub / Play it

Connect

Interested in collaborating? Reach out on LinkedIn.