Selected work / P♠
Projects
Tools, products, simulations, infrastructure, and evidence-gated research—with the honest limits left in.
A model-agnostic, security-first software delivery harness that turns specialist model roles into a gated SDLC with human approvals and inspectable evidence.
Open projectThree public, verifier-first explorations of Conway-99, an Erdős covering problem, and the Krenn–Gu conjecture—publishing exact finite or conditional progress while keeping every global problem explicitly unresolved.
Open projectA deterministic neural-worm terrarium with contact-driven boarding, randomized Seed Forge worlds, causal control lanes, and checksummed replay—plus one unapologetically scripted aerial kickflip.
Open project
A public scientific-computing lab running 302 graded-potential neurons, 95 body-wall muscles, an inextensible low-Reynolds-number body, and a browser viewer as one closed loop.
Open projectA reproducible PufferLib training harness where one compact policy stabilizes and races 2,048 batched Crazyflie-style drones through randomized 3D rings, backed by a measured and hash-checked baseline.
Open projectA living 3D glasshouse aviary backed by a fail-safe weekly asset pipeline that quarantines, validates, optimizes, and provenance-tracks generated bird models before publication.
Open projectA private, production-shaped client value ledger that turns promised outcomes and source-backed proof into monthly briefs, client pulse signals, and transparent renewal readiness.
Open projectAn original Godot 4 vertical slice with an explorable magical hub and a complete, repeatable card duel built around readable enemy intent, mana, wards, status effects, and real win/loss states.
Open projectPrivate, local-first tooling for immutable market-data evidence, deterministic normalization, frozen offline experiments, and reports that keep plumbing checks separate from effectiveness claims.
Open project
An evidence-first Codex plugin and reusable skill that researches current roles and builds truthful, role-specific application packages without quietly inventing a better candidate.
Open project
The deployment plumbing behind this site and its smaller siblings: signed webhooks, health-checked Docker rollouts, and Caddy routing on one VPS.
Open project
An interactive, repository-derived map of the Krenn–Gu proof programme that keeps scoped results, open branches, and failed routes visibly distinct while the global conjecture remains unresolved.
Open project
A small 3D aquarium for the browser, built with React and Three.js and deployed like a real app because apparently I cannot leave anything simple.
Open project
A tiny daily bird site powered by recent eBird observations, a small Express API, and an unreasonable amount of affection for birds.
Open project
A six-person capstone connecting a NestJS/PostgreSQL backend, an Expo mobile app, and a classroom reader service for verified check-ins.
Open project
Two Expo/Firebase prototypes: one for vehicle access approvals and one for QR-based classroom attendance.
Open project2026
Puffer Drone
A reproducible PufferLib training harness where one compact policy stabilizes and races 2,048 batched Crazyflie-style drones through randomized 3D rings, backed by a measured and hash-checked baseline.
My notes
Puffer Drone asks one small policy to learn two related jobs: hold a hover and race through a procedural ten-ring course. The fun part is watching a swarm of drones learn at unreasonable speed. The useful part is everything around the training run that makes the result replayable instead of anecdotal.
The published reference trained from scratch, then evaluated 598 episodes. Its artifact bundle keeps the raw log, effective configuration, checkpoint, report, learning curve, environment record, and hashes together. The headline numbers are visible, but so is the boundary: this is a compact policy in PufferLib's configured simulator, not evidence that a real quadrotor learned to fly.
Puffer Drone wraps PufferLib's native ocean/drone environment in a pinned, containerized, reproducible workflow. Native C physics runs at 500 Hz, the policy acts at 100 Hz, and 2,048 agents are batched for GPU training across hover and randomized ring-race tasks.
The checked-in baseline was trained for 88,604,672 vectorized agent steps and evaluated across 598 episodes. It averaged 8.849 rings per ten-ring race, completed 40.8% of races, and recorded zero hover out-of-bounds events in that configured evaluation distribution.
Validation covers configuration, native vectorization invariants, report generation, source compilation, and artifact checksums. The repository pins its upstream PufferLib revision and publishes the starting configuration, resolved run settings, checkpoint, raw log, derived metrics, learning curve, environment record, and SHA-256 manifest together.