Real projects taken from idea to working code with FeatureBoard — browser games, GPU visualizers, a solid modeling kernel, and a CFD wind tunnel.
The five tagged Live run entirely in your browser — no install, no server. Every project is open source on github.com/valentil. Curious how they're built? → How We Build.
A 3D space runner: ride a dino-astronaut on a comet past photo-textured planets, with a real-time particle tail, starfield, and sun flare.
A 3D typographic toy. Type to spawn physics letters that bounce and stack; words make gravity wells, keys play spatial-audio tones, and animal names summon low-poly zoo animals.
A WebGL wave simulation: particles advance through a 4D tesseract projection via a custom wavelength engine, rendered as speed-scaled helices with interference heatmaps and 10+ live controls.
Two from-scratch canvas views of the Collatz (3n+1) conjecture: an animated "lightning storm" of many sequences racing to 1 at once, plus an interactive binary engine that steps a single number's even/odd trajectory. No charting library — the number-theory and animation engine are hand-written.
A solid modeling kernel written from scratch in JavaScript — and it delivers the operation professional kernels were still fumbling as of 2025: real variable-width edge blends. Tip-weighted rolling-ball blends across any dihedral angle and 3/4/5-edge corners, watertight by construction, exported as standard STEP. Proof: a full blended D&D dice set, machine-checked and rendered with PBR materials in the client-side viewer — try the Benchmarks menu.
A computational fluid dynamics wind tunnel for STL meshes: repair, volume-mesh, solve the flow field, and visualize pressure/velocity and aerodynamic forces. Now imports CAD geometry, auto-assigns boundary conditions, and runs a water-draining benchmark validated to within 0.1% of Torricelli's law.
A working, dependency-free software installer in pure Rust — standard library only, zero crates. It recursively stages a payload directory into a destination, records a manifest of every file it writes, and can verify and cleanly uninstall from that manifest — with automatic rollback if an install fails partway. Ships with a sample payload and a test suite (install/verify/uninstall + rollback).
A GPU-accelerated video montage pipeline with YOLO object-detection-driven auto-editing: multi-clip ingest, color grading, NVENC encoding, and automatic edit decisions. Extracted into a self-contained tool with its own API server — and now bundled with its browser front-end (video_pipeline.html): a clip timeline, encode controls, and live GPU stats.
Our own WebGL 3D Gaussian Splatting viewer — written from scratch, not gsplat.js or a ported renderer. Hand-written PLY parser, spherical-harmonic color, and sorted gaussian-quad rasterizer on top of Three.js. Loads a sample capture on open; drop in your own .ply or generate one with the included training pipeline.
A Chrome extension that uses your browser's on-device Gemini Nano to detect and hide AI-generated marketing slop on LinkedIn, X, Reddit, and Threads. It asks the model for a judgement and gets JSON back — nothing leaves your browser.
A dataset-labeling and classifier-training console. It scrapes frames from video, has Gemini batch-auto-label every image against a class list, lets a human review and correct the verdicts, then trains a YOLO classifier on the approved set and tests it back on video — a 5-step pipeline from footage to model.