batu.

Browser port of the tui-breath terminal app. Static Astro site, plain TypeScript islands, hand-written canvas 2D renderer mirroring the Rust engine and animator layer. Pixel-art channel glyphs (nose / mouth) drawn from in-code bitmaps, no font/image deps. Auto-pause on tab hide, session persistence in localStorage, env-gated AdSense + MailerLite, GDPR consent middleware as a Cloudflare Pages Function. AdSense-funded, no signup, no tracking beyond aggregate analytics.

Astro · TypeScript · Canvas 2D · Cloudflare Pages · Vitest

Terminal breathing guide (v0.2.0). Five-state machine (Menu→Setup→Session→Results→History), pure BreathingEngine (Copy struct), hand-written animation layer (Lerp / Typewriter / Pulse structs — no external dep). 30 FPS async loop, JSON session history. Interactive breath-hold mode with attempt count tracking. Published to crates.io.

Rust · ratatui · crossterm · tokio

Personal Forest-app clone built entirely solo. 4 FastAPI microservices (auth, session, tag, stats) behind Nginx, async PostgreSQL, JWT with auto-refresh and request queueing. React Native/Expo mobile app: offline-first SQLite sync, deterministic timer FSM corrected on foreground resume, forensic session logging (ms-precision), 6-stage SVG tree animation with tag-colour tinting, adaptive push notifications. ~87 commits.

React Native · FastAPI · PostgreSQL · Docker · Expo

End-to-end automated YouTube channel factory running 3 live channels across distinct niches. Telegram clusters, RSS feeds, and a pgvector RAG corpus feed a LangChain research + script pipeline (5 editorial voices); TTS synthesis drives FFmpeg/MoviePy video assembly; a Go upload service handles scheduling and publish. 15+ containerised microservices (Django DRF backend, React SPA). ~600 hours solo.

Python · Django · LangChain · FFmpeg · MoviePy · pgvector · Go · React · Docker

Config-driven, resumable Python pipeline ingesting from Hugging Face, Kaggle, GitHub, direct image URLs, and scraped PDFs into a single HuggingFace Arrow dataset. Deterministic sampler allocates toward a 10M-sample target; state machine, single-instance locks, AWS EC2/S3 execution. ~295 commits.

Python · HuggingFace datasets · Scrapy · pdf2image · AWS

Backend for an AI-powered LinkedIn Chrome extension using Gemini Live. Deterministic state machine, approval-first trust model, immutable action contracts, idempotent execution. ~177 commits.

Python · FastAPI · Gemini Live · Pydantic

Sole backend engineer, first commit to production. 15+ microservices on Google Cloud Run — scraping layer, API connectors, data pipelines, caching, WebSocket streaming, and AI analytics. 80+ live data sources, multi-layer caching (Redis → MongoDB), LangGraph reflection workflows, 5-provider LLM router with retries and backoff.

Python · Node.js · Cloud Run · Redis · MongoDB · LangGraph

Autonomous crypto newsroom — discovers stories, writes articles in 5 distinct editorial voices, and ships them with cover art and audio digests, all without a human in the loop. 3-service pipeline: dual-discovery news fetch (Gemini Grounded + Perplexity Sonar), SHA256 + embedding dedup, 5-persona LangGraph generator gated by a Pydantic contract, Vertex AI imagery + Google TTS studio. Scheduled via Cloud Scheduler.

FastAPI · MongoDB Atlas · LangGraph · Vertex AI

Semi-autonomous smart-contract vulnerability auditor — Solidity, Rust, and C++. Python orchestrating Foundry + Halmos formal verification + Solodit and BlockThreat intel on EVM, with Rust and C++ contract coverage as well. Mermaid diagram output and structured reports.

Python · Foundry · Halmos · Solodit (EVM) · Rust + C++ contract analysis

Two-service Wikipedia API product published on the RapidAPI marketplace. Contract-first OpenAPI design (openapi-generator skeleton + custom impl layer), infobox XPath parser (lxml), async concurrent multi-article fetch via asyncio.gather(). Backend on Google Cloud Run; load-tested with Locust on launch day.

Python · FastAPI · lxml · httpx · Docker · Cloud Run · Locust

Turkish Grand National Assembly observation pipeline reading the daily parliamentary record (tutanak). Scrapy with Puppeteer middleware; ~68 commits. Successor to the Dec 2021 MPs performance-stats scraper — same target, larger scope.

Python · Scrapy · Puppeteer

60K+ Turkish judicial appointments across 13 years, sourced from publicly available HSK government gazettes. Normalized from messy PDF/CSV sources into person-level career histories. Neo4j graph for people/posts/appointments/locations, transition and anomaly analyses. Commissioned by EuroMed Rights for 2 research programmes on judicial independence — the bridge piece between journalism and engineering.

Python · Neo4j · Pandas · PDF parsing

Daily-run Scrapy spiders for three commercial event-ticket platforms. Vividseats build includes a venue-level progress tracker, dual-stream CSV output (events + tickets), and a post-run Google Sheets push via gspread service account. Delivered as paid Upwork contracts.

Python · Scrapy · gspread · pandas

The first coding project after the journalism pivot. Scraper extracting question proposals, research proposals, parliamentary questions, and MP resumes from the Turkish parliament website, plus an analytics layer calculating performance metrics for MPs and political parties. Later commercialized.

Python · Scrapy