PROJECTS
guest@rrs:~/projects$
SUN-EARTH-MOON SYSTEM VISUALIZER
My kids keep asking the same two questions, and both are weirdly hard to answer with words: why is it always the same face of the Moon we see, and how can the Sun and Moon be up at the same time? So I built a thing you can poke at instead. It's the rolling-coin analogy made interactive -- a tidally-locked Moon you can unlock, real moon phases, and an observer's sky that shows exactly when both the Sun and Moon clear the horizon. Periods and phase geometry are faithful; distances are honestly not to scale.
SOLAR SYSTEM
Ever wonder where Neptune is right now? This renderer uses real orbital data from NASA's JPL Horizons system to draw an accurate snapshot of the solar system on any date from 2001 to 2031. Check a birthday, watch the planets race each other in timelapse, or filter down to just the moons. Any frame you land on saves out as a wallpaper.
SATURN V ASCII LAUNCH
Started because I wanted a rocket launch playing every time I opened my terminal. A fastfetch splash screen turned into a full truecolor ASCII rendering engine -- the Apollo 11 liftoff, frame by frame, with real flame physics and smoke that dissipates top-to-bottom. The full NASA footage streams as 2MB chunks of colored characters. Install instructions are in the repo.
PERSISTENT MEMORY FOR LLM AGENTS
LLM agents forget everything between sessions. I built aelfrice to model memory as scored beliefs with Bayesian confidence tracking, detect corrections without LLM calls (92% accuracy), and bridge vocabulary gaps in retrieval (99.5% recovery). No embeddings anywhere -- FTS5 keyword search plus typed knowledge graphs. Benchmarked in public, with the re-runs published even where they came back lower. Now at v4.0: temporal-spine retrieval, which lifted coverage 14.6pp on a fresh LoCoMo sample, plus junk demotion.
DIFFERENTIABLE ROCKET SCIENCE
In 2021 I tried to train an RL agent to fly to the Moon for a grad astrodynamics course, and it never learned a thing. In 2026 I went back, found the real bugs, and rebuilt it on a better idea: backpropagate the fuel bill through the physics itself. The result circularizes 92.3% of random orbits on the full 3-D task, and spends 1.03x the analytic-optimal fuel at its best-fuel checkpoint. It also rediscovers a real operational J2 maneuver from the raw objective, and pulls off a verified ballistic lunar capture with no capture burn. The 2021 failure ships in the repo, unmodified.
MINING SATURN'S RINGS
A mission concept for trawling multi-tonne blocks of 99.7%-pure water ice out of Saturn's B ring and flying them to Earth orbit -- the cargo doubling as the return propellant tank. I ran 215 pre-registered analysis rounds trying to kill it, and in the end they did. Under closed-loop mass accounting, zero of 1,744 mission paths complete the round trip, the one paper corridor left stacks three unproven technologies, and the financial model prices its water at three times the best defensible market. The retractions ship in the repo alongside the claims.