Resource bay
A lightweight directory for GPT assistants, code experiments, and useful public links.
resource.project_index
Every public repository currently visible from the GitHub profile, with private repos excluded from the public site.
Live Photo editor for iPhone that saves genuine Live Photos back to the library, plus a full browser edition on this site: GPU-graded looks and adjustments, Loop, Bounce and Long Exposure, a soundtrack timeline with voiceover, crop, and MP4, GIF or key-photo export.
Anime quant research club (v4): six students propose, backtest and audit trading ideas with an LLM brain, then paper-trade the survivors. A Python engine with point-in-time S&P 500 data, next-open fills and real costs; rarity is the empirical-Bayes expected live Sharpe, behind deflated-Sharpe gates over every pull and a sealed lockbox. Chibi club room, visual-novel storyboard, all-new art.
Android phone assistant that takes calls under per-contact rules, says up front that it is an AI, reads key details back, and turns requests into tasks for a cloud model or Codex. With a privileged install it also answers by voice, turns abuse on calls you take into calm captions, or speaks replies you pick; on a stock phone it handles requests over SMS. Nothing goes out until you approve the draft with a fingerprint, and every event lands in a hash-chained archive on the phone.
Native macOS pinyin input method: a 1.44M-entry dictionary engine decodes every keystroke in milliseconds, and a local Qwen3-0.6B re-ranks the candidates using the text before the cursor, but only when a calibrated gate expects the dictionary to be wrong. On a held-out corpus it lifts accuracy from 37.8% (dictionary only) to 46.1%, and it matches Apple Pinyin once there is context. Learning is AES-GCM encrypted on the Mac, and the app has no network access at all.
Original papercraft strategy card game about job offers, now in 2.3 public beta: your offers and degrees become cards, dragged onto a four-slot table with animal allies, while two era changes rewrite the rules mid-match. Bot matches at four difficulty levels and the tutorials run locally in the browser; friend rooms, accounts and match history run on a Cloudflare Worker with Durable Objects.
LLM mission architect for Kerbal Space Program 1: Claude Code or Codex, called through the local CLI with no API key, breaks one plain-English goal into atomic flight primitives with calculated parameters, which are then flown live over kRPC and a C# MechJeb bridge. Every rocket is sized from physics and must pass a Codex review of its three-view drawing before launch. Flown so far: a crewed Mun landing and return, a crew landed on Duna, and relay constellations at Eve and Duna.
Lean 4 results around Erdős problem 488 on primitive sets. It proves the original density inequality for every primitive generator set with excess at most 15, shows that 16 is exactly where the auxiliary estimate first fails, and refutes two conjectures from a March 2026 manuscript with explicit counterexamples. The proofs, found with a Codex multi-agent workflow, use no sorry or native_decide; the full problem remains open.
World-model driving stack for Forza Horizon 6 in Offline Solo / Free Roam. It connects Windows screen capture, Forza Data Out UDP telemetry, route features, a DreamerV3 actor-critic, virtual Xbox controller output, crash recovery, replay storage, and stress tests under a tight live decision budget.
Adversarial AI referee for empirical economics. It reads a paper, identifies hidden specification choices, researches unfamiliar methods, looks for public data, re-runs headline results across defensible alternatives, and reports fragility, p-hacking, cherry-picking, and replication risks in plain language.
Local-first research assistant that turns workday traces into private notes, idea candidates, daily briefings, and a searchable research backlog. The project includes passive observation, novelty and feasibility judging, multi-provider model routing, optional screenpipe/OCR capture, RAG context, and a staged research pipeline.
Chinese dating-chat strategist packaged as an agent skill. It reads chat or social screenshots, estimates interest and risk, drafts low-oiliness replies, separates user-facing messages from private advice, keeps multi-person memory, and includes stop-loss style rules for ambiguous or manipulative interactions.
Reusable Claude Code workflow for economics and economic-history papers. It indexes a local corpus, forces direct source reading, builds evidence cards and claim maps, plans sections, audits citations, runs field reviews against rubrics, assembles drafts, renders PDF, and checks page-range gates.
Claude Code methodology skill for structural industrial organization research. It covers BLP/random-coefficients demand, nested logit, instruments, weak identification, Monte Carlo validation, DiD/RDD, market definition, counterfactuals, merger simulation, replication practice, and paper-writing checks.
Browser extension that screens Zhihu authors before their posts reach you: it samples the current card plus the author's three latest posts and asks a local OpenAI-compatible model (Ollama, LM Studio, vLLM) whether they write flame-war bait. Matches are folded into a quiet placeholder, hidden or blurred on the page, with allow and block lists, thresholds and a week-long cache. The prompt judges text behaviour only, never identity, and posts pass through when the model is down.
Unofficial Edge extension for China Datang's recruitment site, which shows parts of its Chinese text as mojibake. It detects mis-decoded UTF-8 byte runs, restores only text it can decode unambiguously (including text the page adds later), and can undo the fix per page. It runs on that one site, stores nothing but an auto-fix switch, and makes no network requests.
梗王大乱斗 · Meme Kings Smash: a Smash-style browser platform fighter starring seven Chinese internet meme kings. It has percent-based knockback with DI, shields and parries, dodges, grabs, ledges and techs, a 斩杀线 kill line marked on every health bar, break-to-claim item boxes and a floating Final-Smash badge, across six stages with switchable hazards. One to four players on keyboard, gamepad or touch take on level 1–9 CPUs, with a scrolling danmaku crowd, a Persona-style red/black UI and a soundtrack synthesized entirely in WebAudio.
Automation script for discovering United Nations internship listings. It uses Selenium with stealth settings, extracts job metadata and links, filters already-applied Job IDs from an Excel tracker, geocodes duty stations, organizes results by country, and writes a workbook for deadline tracking.
Small Windows GUI utility that keeps selected programs pinned above other windows. The repo contains the built Qt application artifacts, supporting DLLs, and a concise README describing support for one, two, or three top-locked windows.
Single-file prototype that turns player names into ASCII-derived attack, defense, health, and battle power, then runs a turn-based auto battle with randomized skills and colored terminal output. It reads like an early experiment in playful procedural rules.
Fork of wflixu/RClick that ends the macOS 15+ “access data from other apps” prompt, which returned after every Finder restart, by moving the Finder extension off a non-team-prefixed app-group suite. The main app is replaced by a windowless, event-driven LaunchAgent that sends HMAC-signed menu config and runs every action of the Windows-style right-click menu, using 0.28 s of CPU over 52 minutes.
Fork of musistudio/claude-code-router that adds an account-pool scheduler for Claude Code. It estimates each request's tokens before it leaves the machine, sends it to the first authorized account with enough local budget (ordered-fit), and reconciles the reservation against real usage. It only manages local budgets; it does not get around provider-side limits.
Forked writing skill for removing common signs of AI-generated prose. The README describes 29 detectable patterns, before/after rewrites, voice calibration from writing samples, and a second-pass audit for text that still sounds synthetic.
Forked experimental skill generator that turns chat logs, screenshots, photos, and subjective notes into a remembered-persona AI skill. It includes appendable memory, correction layers, version archives, rollback support, and explicit caution notes about emotional overuse.
The public lab site itself. This redesign turns the old static blog output into a dark Visual-Studio-Code-like interface with image-generated hero art, local search, archives, resources, AboutMe, and this GitHub project registry.
assistant.links
A GPT assistant focused on learning the logic behind programming problems before jumping to solutions.
Open GPTA writing assistant for refining statements with clearer structure, voice, and evidence.
Open GPTA focused advisor for IELTS writing practice and feedback loops.
Open GPTSource code, experiments, and public project history.
Open GitHubA formal CV attachment for applications and academic context. The site keeps the public alias; the document can carry formal details.
Open CV