Type a vibe like “dark academia but rain-proof” and get three cohesive, rentable outfits, presented by a loving parody of Cher’s closet computer.
The interesting problem is coherence: pieces that look good together, not just pieces that individually match the vibe. That part is deliberately not an LLM. A model call turns the vibe into structured attributes; hand-written stylist logic assembles the outfits from Nuuly’s catalog, and every piece links to its live product page.
Replay one aircraft’s entire day on a seatback-style map, with real ADS-B pings, honest interpolation, and the gaps where no receiver heard it.
Opens on a recorded Delta 1 flight-day (JFK → Heathrow → Salt Lake City) from OpenSky Network data and starts replaying immediately. The playback is honest about its data: positions interpolate only between real pings, and coverage gaps are drawn as gaps rather than smoothed over. Live search of arbitrary aircraft works when run locally with OpenSky credentials; the hosted demo ships the saved day because OpenSky blocks cloud IPs.
Plant chords in a garden and feed it music to make it grow: a from-scratch Web Audio synth plus a live music-theory tension trellis.
Every chord you play sows a plant and lights up a tension trellis, a circle-of-fifths ring where each chord’s pull toward the next is drawn as a vine. Plant a G7 and the vine to C throbs until you satisfy it; resolve, and the tonic’s plant bursts into bloom. Resolving progressions feeds the garden, wandering leaves seedlings, and different chord qualities grow different species. All sound is synthesized from scratch with the Web Audio API. No goal, no score: a toy that secretly teaches functional harmony.
ML day-trading signals for a stock watchlist: walk-forward gradient boosting, a backtester, and this static dashboard of the results.
The Python engine computes ~15 technical indicators and trains a walk-forward gradient-boosting model per ticker, so the model never sees the future it predicts. The backtester replays out-of-sample predictions against buy-and-hold (total return, Sharpe, max drawdown, win rate), and the dashboard shows today’s signals with charts and equity curves. Research tool, not financial advice; it says so itself.
Thousands of glowing particles under hand-written Verlet physics. No framework, no physics library.
The physics is the project: Verlet integration at a fixed 120 Hz timestep with render-time interpolation, a spatial hash grid over flat typed arrays so particle–particle collisions stay tractable at 20k particles, and zero allocation in the hot loop. Attract and repel with mouse or finger; tune gravity, bounciness, and trails live.
A choose-your-own-adventure played beside a live view of its own state machine. Your path lights up as you move through the story.
A haunted-lighthouse mystery with 22 nodes, six endings, and one choice that only exists if you picked something up earlier. The second panel is the story’s actual state graph, not an illustration of it: the current node and every traveled edge light up live, and Undo un-lights them. The state machine is the visualization.
Paste any text, pick a vibe (corporate, Shakespearean, passive-aggressive HR), and watch the rewrite stream in word by word.
A single serverless function receives the text and style, calls the Anthropic API, and streams the rewrite back over server-sent events. The API key never ships to the client. Each style is a system prompt plus one or two few-shot examples, which do the heavy lifting for voice consistency. The streaming is most of the fun.
Research
A cyber attack modeled as a game between two agents with minds. Red, the attacker, works a reconnaissance-to-compromise kill chain, but can be endowed with human decision biases like base-rate neglect, loss aversion, and the sunk-cost fallacy. Blue, the defender, watches Red act and performs Bayesian inference over hypotheses about which bias is distorting Red’s choices, turning an attacker’s psychology into a defensive signal.