Nabil Babaslimane

One 805-line HTML file. No engine, no framework, no build step, no dependencies.

Instead of carrying light as three numbers, Carat carries a wavelength for each ray and recomputes the stone's refractive index at every surface it touches, so dispersion falls out of the physics rather than being authored afterwards. The 65-facet brilliant is modelled as 65 intersected half-spaces, not a mesh, which turns intersection into one ordered loop over planes.

It holds a 36 ms median frame at 912×638 on a 2-core Ryzen 3 APU with integrated graphics. I then wrote a second, independent model of the optics in numpy and validated rendered output against it. It found nine bugs, every one of which had produced a plausible number from a broken renderer.

• Case study ↗
• Live demo ↗
• Source ↗

An agent built on top of OpenClaw, an open-source assistant that runs locally, plus the Python command line tool underneath it. The hard part was not the searching. It was writing down, before anything ran, what the agent may and may not do on my behalf: draft-only while eligibility was unresolved, no CAPTCHA solving, no scraping of sites that forbid it, a daily cap, and a ledger entry for every attempt.

The tool is standard library only and keeps one SQLite file. It reads five job boards, matches whole words so that intern does not drag in internal, applies two filter groups that must both be satisfied, and renders digests short enough to read on a phone. CI runs ruff and pytest.

• Source ↗
• Companion project: Replay ↗

Job adverts change or disappear, so a saved link is not always enough. Replay lets me capture a listing as I found it, capture it again later, and review a timeline of what changed.

It complements career-ops: Replay can import discovered postings from its SQLite database in read-only mode, then keeps its own snapshots. It is a one-time import, not live sync, and does not import application statuses or private notes.

• Source ↗
• Case study ↗
• Related project: career-ops ↗

A tool that follows predictors on Polymarket and ranks them by win rate, so attention goes to the ones with a record rather than the ones with an opinion. It reads their recent activity, uses a language model to propose emerging markets worth watching, and writes the daily picks to a running log for review.