I'm Baran Dursun.
I build and ship data-heavy products.

Fragrance enthusiasts track their collections across scattered notes and spreadsheets, with no single place to identify a bottle, log what they wear, or see what friends are wearing.

Solo end to end: a 58-endpoint Node.js/Express REST backend with Redis caching behind a React Native (Expo) client — ~18K lines of TypeScript. Barcode scanning and free-text lookup run on PostgreSQL trigram (pg_trgm) search with a custom confidence-ranking layer over 63,751 fragrances across 1,772 brands and 1,861 scent notes; a content-based recommender ranks by IDF-weighted scent-note overlap scaled by brand-tier affinity.

Live on the App Store, shipped through CI with 528 automated tests and a schema-drift guard. A self-initiated security audit of the app in production uncovered a PII exposure in its auto-generated REST API — fixed at the Postgres grant/RLS layer with a regression test. Also ships B2B analytics endpoints designed for fragrance houses.

Coming back to a series after a break, you want the story so far — but every recap, wiki, and search result is a minefield of spoilers for episodes you haven't reached yet.

TypeScript end to end: a Hono + PostgreSQL (drizzle-orm) API behind a React 19 web app and a Chrome MV3 extension — roughly 6,600 lines with 184 automated tests. Plot events are extracted on demand from episode synopses with the Claude API and streamed back as catch-up recaps (three lengths), single-episode summaries, and spoiler-bounded character cards. The extension's content scripts read your current episode straight from the Netflix, Prime Video, Disney+, and Max players.

Live at recapengine.com behind Google sign-in, deployed on Railway. The interesting constraint: spoiler-safety is a database invariant, not a filter bolted on afterward. Every episode carries a contiguous order number, and a single query — the only path from stored events into any prompt — clips everything past your position, so leaking a future event is structurally impossible. Episodes ahead of you are never even read until a later boundary needs them.

Canadian macro data is scattered across Statistics Canada and the Bank of Canada in incompatible formats, making it slow to combine indicators or reason about scenarios.

Ingests 9 indicators into a PostgreSQL medallion (bronze / silver / gold) warehouse, then runs VAR + XGBoost forecasting with Monte Carlo scenario bands behind a FastAPI + React dashboard. The interesting part was layering RAG natural-language Q&A over the same warehouse the numeric models read from.

Live, serving a composite health score, scenario forecasts, natural-language Q&A, and one-click PDF reports.

Feeds and news sites optimize for engagement, burying the signal under clickbait and the same story duplicated across sources.

A scheduled pipeline polls Finnhub, RSS feeds, and Hacker News, scores every article for relevance and clickbait with the Claude API, and renders a dense three-column broadsheet from Postgres. The whole thing runs as a single Railway service.

Live on Railway, regenerating the front page on a schedule.