Projects

AI-chat-driven personal task organizer for a couple, with task board and calendar views.

Personal daily-briefing app aggregating weather, news, sports, stocks, crypto, and more.

Daily briefing platform for Asian dramas, stars, music, and reality TV news.

Video-first language-learning platform teaching Japanese/English through real YouTube lessons.

Personal finance-news dashboard with market, earnings/SEC, and AI/tech news plus a watchlist.

Turns pasted notes into flashcards, adaptive quizzes, practice tests, and an AI tutor.

3D Tokyo life-sim for learning Japanese, with AI-driven NPC conversations.

Ten interactive browser ports of an AI curriculum, from classical AI to agents, every algorithm written from scratch.

Educational swing-trading research platform and paper-trading coach, built with Streamlit.

Travel command center - world map, couples journal, photo memory boards, and AI trip planning via Yuu.

Food exploration OS - interactive restaurant map, food journal, photo scrapbook, and an AI concierge.

Private fork of OpenChatCut, an AI-assisted, transcript-driven video editor.

Personal AI command system built on OpenJarvis for tracking concurrent Claude Code sessions.

Hang an artist's work in 141 real photographed rooms for auction listings, with a video compositor mode.

Real-time AI performance camera editor.

Turn a screenshot into a polished animated product showcase, right in the browser.

Turns a one-sentence idea into a production-grade coding-agent prompt: architecture, schema, APIs, tests, and a build plan.

Data-driven player cards generated from GitHub, Spotify, Discord, Instagram, and more.

Short-term fashion rental - rent everyday clothing for tonight, the weekend, or the week.

Chrome extension that turns any tab into a research workspace.

Personal AI agent for task triage, decision support, and a daily digest - a terminal that learned graphic design.

Webcam tool that opens a McDonald's job application whenever you look at your phone.

Single-file serverless URL shortener.

Pixel-art habit tracker where your habits grow a garden.

Personal command center / Life OS unifying school, career, research, and personal life into one ranked priority view.

SWE/quant/ML learning platform with dozens of hands-on skills and a sandboxed code runner.

Persistent virtual hangout room for friend groups with voice/video, camera bubbles, and drawing.

Read-only paper-trading dashboard mirroring an EMA-crossover strategy on live market data.

Password-gated scanner comparing sportsbook odds vs. prediction markets to flag +EV bets.

Prediction-market (Kalshi/Polymarket) +EV scanner and paper-trading terminal for sports markets.

Interactive global job salary, career-transition, education-impact, and industry-trend tracker.

Tracker for an entire grad-school and job-application pipeline, from prospects to offers.

Private app for a couple to connect Oura Ring accounts and share sleep/health data.

Group members capture daily moments that auto-assemble into a shared vlog at a cutoff time.

Screen-scraping poker assistant for friendly, no-money home games.

Local, transparent Monero/RandomX mining laboratory and dashboard for Apple Silicon Macs.

Private fork of an open-source real-time global intelligence dashboard - news, geopolitical, and infrastructure monitoring.

Personalized fork of Project AIRI, an AI companion / virtual character app.

Local-first session tracker for concurrent Claude Code sessions across projects, terminals, and VS Code windows.

3D interactive bookshelf portfolio site.

Tiny single-page site that sends a message by email or Discord DM via a shared access code.

Webcam-based app using on-device ML to detect phone use and trigger a fullscreen alarm.

3D generative dungeon RPG where every floor has a reason to exist.

3D living-world anime-fantasy RPG with an original trapped-in-a-VR-MMO premise.

3D generative open-world fantasy RPG.

Living anime social-life RPG set in a Japanese coastal university town.

Stand in Edo, 1600 - an inhabited 3D historical world, not a museum.

3D alternate-history engine - change a turning point and watch the world diverge.

3D generational civilization sim - a colony ship, 120 persistent colonists, and emergent history.

Persistent 3D artificial-civilization simulation with emergent citizens.

Artificial-society experiment - 200 autonomous minds, one persistent world, no script.

Detective game where the murders emerge from simulation instead of being scripted.

Butterfly-effect engine that propagates one small change through a causal world.

Live the life you didn't choose - a fictional plausibility simulator.

Multi-sport fantasy league app - draft and compete against 7 AI managers with distinct personas.

Fast browser-based 3D 1v1 build-and-shoot duel game against a bot or a friend.

Real-time multiplayer shotgun-duel game inspired by Buckshot Roulette, with a Career Mode.

Original from-scratch online chess platform with 10 Stockfish bot personalities and variants.

Social multiplayer card game hub - Hold'em, Blackjack, Baccarat, Dou Dizhu, and Old Maid, with friends or AI.

Multiplayer social-deduction game in the spirit of Among Us, with full roles, maps, tasks, and AI players.

Real-time multiplayer music-guessing party game - buzz in, name the song, steal points.

Browser-based multiplayer mini-game platform - no download or account needed, just a room code.

Twelve original daily logic, word, pattern, and memory puzzles - play solo, race a friend, or run a Speed Round.

Story-driven 2D anime life simulator blending visual novel, social-strategy, and turn-based RPG.

Anime-psychological RPG with Undertale-style bullet-dodging combat and visual-novel dialogue.

Tiny-planet exploration game - walk a curved world and carry hand-written messages between NPCs, with realtime multiplayer presence.

Experimental desktop environment where apps, files, tasks, people, and context behave as one.

Social network built around discovery and relationships rather than engagement.

Persistent rooms for the people you love - leave a note, pin a song, come back to what changed.

Realtime messaging and community platform with conversation memory.

Desktop-class email client.

Local-first knowledge environment - block editor, backlinks, canvas, graph, timeline, and study mode.

File manager that understands relationships, projects, versions, and meaning.

Simulated developer terminal with tabs, panes, command blocks, failure intel, and a command composer.

Evidence-first research workstation - question, plan, evidence, claims, contradictions, confidence, report.

Conversational Monte Carlo simulation sandbox.

Reproducible AI evaluation lab - what each model is good at, how reliable, how fast, and whether a new version regressed.

Controlled lab for testing, benchmarking, replaying, and debugging AI agents that operate web interfaces.

Design intelligence platform - design contracts, component vault, motion, typography, and anti-pattern audits.

Cross-repository structural code index and search across the whole portfolio.

Reverse-mode autodiff deep-learning framework built from scratch on NumPy, verified against PyTorch.

Decoder-only Transformer language model built from PyTorch tensor primitives - forward pass, training loop, and sampler.

Multi-agent orchestration framework in raw Python - typed messages, budget reservations, replayable traces, no LangChain.

Robot autonomy stack from the math up - mapping, Monte Carlo localization, A*/RRT planning, and exploration, no ROS.

An LLM agent that drives a simulated robot through a typed tool boundary.

Language-conditioned pick-and-place in MuJoCo - perception, calibration, inverse kinematics, and closed-loop verification.

Offline retrieval-then-ranking framework for recommendation and search, with leakage enforced as a runtime invariant.

Open-source template taking a tabular model from training to a validated production service.

ML inference benchmarking toolkit built around a dynamic batching scheduler and standard backend adapters.

LLM evaluation toolkit where judges are instruments that must be calibrated and CI gates fail loudly.

Model-training toolkit that makes data leakage structurally hard rather than merely discouraged.

Correctness-focused quantitative finance library for Python with lookahead-safe historical APIs.

Audits tabular datasets for data-quality defects and ML leakage - 32 detectors, each deterministic or heuristic.

Open-source Python library for detecting drift between two windows of data.

Why decoding one token from a 7B model means streaming every weight out of RAM.

Nine browser demos from minimax to agents, every algorithm implemented from scratch.

Sentiment, zero-shot classification, and fill-mask from three pretrained models - zero training.

Keras autoencoder that squeezes 1,797 handwritten digits into 2 numbers each.

Character-level neural name generator in one PyTorch file.

50,000-row house-price benchmark testing LightGBM's speed claims with a stopwatch.

What spaCy pulls out of a news paragraph - entities, structure, and timing from a real run.

Cubic curve fit in raw TensorFlow GradientTape, with the Keras removed.

Benchmark of LLM agents writing Lean 4 proofs, where a proof only counts if Lean verifies it.

Classical vs. neural models forecasting mouse visual-cortex activity under strict future-time and cross-animal holdouts.

Benchmark of generative-design methods for multi-objective middle-housing layouts, with Pareto analysis.

Browser psychophysics platform and estimator benchmark for serial dependence in perception (simulation-only).

Measures how much real signal machine-learned systematics correction erases, with a DESI cosmology case study.

Are AI-generated radiology research syntheses grounded in report evidence and published literature?

Do AI-generated statistics exam questions actually measure knowledge? A validity benchmark.

Forward-in-time benchmark of classical vs. machine-learned aftershock models on fourteen real earthquake sequences.

How much seasonal-climate variability orbital forcing explains, and how robust that is across models and scales.

How robust are NLP measurements of changing gender representation in U.S. news, 1900-1963?

How reliably retrieval and RAG systems find, combine, and cite sparse evidence for rare-disease questions.

Benchmark for ML restoration of degraded historical audio: does cleaner mean closer to the original?

Benchmark of evidence-grounded LLM agents for biological literature synthesis.

Reproducible large-scale-structure clustering and ML systematics-correction study on public DESI DR1 data.

How AI-generated product framing shapes consumer belief, choice, and market outcomes.

Benchmark of self-supervised and unsupervised representation learning for biomedical images.

Computational mapping of cultural change and narrative structure in the Journal of American Folklore, 1888-1930.

Can Sentinel-2 satellite imagery detect disturbance signatures of river sand extraction?

Measuring urban infrastructure from street-level imagery in Indian cities - and how biased uneven coverage makes it.

Measuring how suspense evolves across 92 public-domain short stories, and whether language models reproduce it.

Audits how reliably NLP reconstructs the social networks of the Icelandic family sagas.

Compares random, greedy, annealing, NSGA-II, and constraint search for generating building floor plans.

Can LLM agents execute reproducible scientific data-analysis workflows? Audited on high-energy physics.

Pre-registered repeated-decision experiment on when people should rely on AI recommendations.

Benchmarking long-horizon coherence - same character, same world - in generative visual narratives.

Longitudinal, evidence-backed panel of commercial AI products - pricing, features, and providers over time.

Browser-based behavioural experiment and research platform on slow-change blindness.

Welfare simulator for personalized pricing, with Monte Carlo experiments and econometric recovery.

LoRA domain specialists evaluated against their own base model under identical conditions.

Why deep networks fail to train - 132 instrumented runs on gradient propagation in deep MLPs.

Does semantic change in earnings calls predict volatility? An NLP x finance study.

How Fed policy surprises propagate across markets - an FOMC event study.

When extreme stock selloffs reverse, and when they keep falling - 14,678 events studied.

Do extreme closing price moves persist overnight? An intraday event study.

What happens right across the border after a minimum-wage increase? A border-county difference-in-differences study.

Tail-risk study of NYC taxi travel-time reliability across 75M cleaned trips.

When estimation error beats optimisation - a Monte Carlo study of mean-variance portfolios.

Survival analysis and trajectory clustering of how customers approach churn.

When forecasting complexity loses to a naive baseline.