I turn ideas into products people use.

Students copy-paste PDFs and YouTube transcripts into chatbots and get unstructured walls of text. The information is there — the learning experience isn't.

Designed an ingestion pipeline that normalises mixed media (PDFs, video transcripts) into a single content graph, then orchestrates LLM calls to produce summaries, flashcards, and quizzes with predictable shape.

• Keeping AI output deterministic enough to render in fixed UI components
• Handling long-context documents without blowing token budgets
• Reliable extraction across malformed PDFs and auto-generated captions

Next.js · Edge functions for ingestion · Background workers for embedding · Structured output via JSON schema · Postgres + pgvector

Students could move from raw source material to a study session in under a minute, with consistent interactions across every generation.

Integrating AI is relatively straightforward. Creating predictable user experiences around unpredictable AI output is significantly harder.

Bookmark managers are graveyards. Links get saved, titles stay cryptic, and nothing is actually findable a week later.

Built a per-user vault where every saved link is enriched in the background — summary, tags, and topic clusters generated automatically. Search works on what the page is about, not just its title.

• Designing user-scoped data with row-level security from day one
• Reliability when third-party pages block scrapers or change shape
• Keeping metadata fresh without re-processing the entire library

React + Vite · Supabase auth & Postgres · Edge functions for enrichment · Queue with retries · LLM for metadata + classification

Recovering a half-remembered link became a question of intent rather than recall.

Small workflow improvements often create more value than large feature sets.