Most of these exist because I needed something and the available options were either overkill or didn't exist.
Nine financial tools in a single HTML file: live FX with 30-day sparklines, crypto prices, US stock quotes, real inflation data by country, DCF valuation, compound interest, loan/mortgage, stock return, and purchasing power calculators. Plus an ASML tracker that plots the live price against my own DCF range.
No sign-up, no backend, no tracking. Four live APIs (ECB, CoinGecko, Alpha Vantage, World Bank), all client-side. The kind of financial toolkit I wanted as a student reading annual reports for the first time.
Browser-based video labeling tool. Frame-by-frame navigation, timestamped multi-class annotations, CSV export. Built for the Tsinghua lab when CVAT and Labelbox turned out to be absurdly heavy for what we needed: timestamps and labels, not bounding boxes.
End-to-end ML pipeline that recommends music based on detected mood — speech recognition → sentiment analysis → content-based filtering → local playback. The constraint: entirely offline, zero API calls, everything runs on one machine. Not production-grade, but a useful exercise in building without a safety net.
C++ CSV parser. Exists because pandas choked on an 8GB file during a research task. Memory-mapped I/O, zero-copy parsing. Finishes in ~40 seconds what pandas does in ~3 minutes. Small project, but apparently "I wrote a C++ parser because Python was too slow" makes for a good interview story.
Interactive visualisation of AI's environmental cost — energy, water, e-waste — built for a Leiden interaction design course. Users compare models via Chart.js dashboards and a Three.js 3D landscape that degrades as you increase AI usage. Melodramatic, but it got the point across.
This site. Static HTML, GitHub Pages, custom domain. No frameworks, no build step, no dependencies. Fourteen files, 189KB total, loads in under a second anywhere. I understand every line of code because I wrote every line of code.