Demos you can play with, textbooks I wrote around Dartmouth courses, extending Professor Peter Chin’s lectures with derivations and worked examples, and the theses and papers behind the work.
Demos from the research, playable here. The live pages are the ones published with each repository and load when you ask.
Do “regime-switching” strategies actually work? Tested the way an investor would have lived them since 1926, the published gains vanish, and the best recent method fails outside its sample. The fix is planning: a cube-root law for when acting on an uncertain belief is worth the cost. Drag the cost slider and watch naive switching collapse while the planned version survives.
Melody models are graded on tunes they supposedly never saw. A transposition-proof search over eight corpora finds that a third of the JSB Chorales test set repeats a training melody, and that the leak flips which model looks best. Play each test tune against its twin, and hear what a model trained on duplicated data copies.
Sliders for generative AI music. Each one is a small LoRA on a text-to-music diffusion model that nudges a single quality of a piece (happier, groovier, more jazz) and is graded by measuring the audio that comes out, so a slider has to do what its label says. Weights are open on Hugging Face.
“He is called as you know the apostle of the indies.” Whisper reads the recovery word for word.
“Privacy-preserving” speech recognition can leak the audio itself. Federated learning sends model updates instead of recordings; one update is enough to rebuild the speaker’s words exactly, with no transcript and no guesswork. Switch between the original and the recovery mid-sentence.
Cleans up and masters a finished song on its own. It measures what is wrong before touching anything: exact signal processing for clipping and echo, a neural network for room reverb, and the norms of 103,838 released tracks for tone and loudness. The explorer lets you set each reading and watch the decision.
Taught a Go engine when to stop thinking. Strong players glance at easy moves and sit over hard ones; engines spend the same effort on every move. Mikiri wraps KataGo, the strongest open AlphaZero-style engine, with a learned stopping rule for its Monte Carlo tree search and a retrieval memory of past searches. Same compute, +228 Elo over 1,000 games.
Asked how much the famous “mispricings” in prediction markets were really worth. Rebuilt 23.7 million trades across a million resolved Polymarket markets, then stress-tested every claimed edge with real fills, fees, stale data, walk-forward validation, and LLM forecasters. Most of it disappears; two effects survive, and one is fading.
Written around Dartmouth courses, with derivations and worked examples.
Mathematical Foundations of Machine LearningENGS 96, DartmouthThe mathematics a machine-learning course assumes you already know, written so you actually do: linear algebra, calculus, probability, and optimization built up from first principles.Read PDF 
Game-Theoretic Design, Learning, and EngineeringENGS 102, DartmouthGames, equilibria, mechanism design, and learning in strategic settings, from matrix games through multi-agent learning.Read PDF 
Principles of Machine LearningENGS 106, DartmouthThe graduate ML core: estimation, regression, kernels, neural networks, unsupervised methods, and the theory that connects them. I later supported the course as Head TA.Read PDF 
High-Dimensional Sensing and LearningENGS 109, DartmouthCompressed sensing, sparse recovery, and low-rank models: how structure makes high-dimensional problems tractable.Read PDF Three theses across engineering and music.
MODULO: A Generative Audio Workstation for AI-Native Music Co-CreationM.S. Engineering SciencesWhat a DAW becomes when generation is native to it. Five user studies with working musicians.Read PDF 
Tokenized Curriculum Learning for Multi-Effect Neural Restoration of Full-Mix MusicHonors thesis, Engineering SciencesA 1.08B-parameter token-domain U-Net trained with a staged curriculum to restore five types of degradation in full-mix music.Read PDF 
My Life in SoundHonors thesis, MusicThe writing behind my music honors thesis: a multi-genre album I composed, performed, recorded, and mastered.Read PDF Research writing across audio restoration, generative audio, speech privacy, search, markets, and agents. Newest first; each links to its code where there is some. The newest are manuscripts in preparation, not yet peer reviewed.
Planning Through Regimes: A Century of Real-Time Evidence and a Cube-Root Law for Regime Timing Under Transaction Costssolo · in preparation for ACM ICAIF 2027Regime-switching allocation tested in strict real time since 1926. The published edges come from look-ahead; the state-of-the-art jump model fails outside its sample; and a POMDP with holdings in the state gives a no-trade band in belief space with a cube-root width law that keeps regime signals alive under real trading costs.Read PDF Code 
Transposed Twins: Benchmark Leakage and Memorization in Symbolic Melody Modelssolo · in preparation for TISMIRA transposition-invariant audit of eight melody corpora. 36% of the JSB Chorales test set repeats a training soprano and half of a random PDMX split would; retraining without the twins shows the leak rewards copy-prone models enough to reverse a ranking, and grows with transformer size.Read PDF Code 
Lacquer: Deciding What to Fix in a Finished Mixsolo · in preparation for DAFx 2027Automatic restoration and mastering that measures before it acts: sparse declipping, cepstral echo removal, a band-split transformer for reverb, and tone, image, and dynamics held within the measured range of 103,838 released tracks.Read PDF Code 
Frames in the Span: Exact Recovery of Long-Form Speech from Federated ASR Updatessolo · in preparation for ESANN 2027One federated update from a speech recognizer gives back the client's audio features in closed form: 98.5% of 1,417 LibriSpeech utterances up to 6.2 s, sequential decoding to 35 s, no transcript and no optimisation.Read PDF Code 
Mikiri: Learned Stopping and Exact Search Memory on a Frozen Go Enginesolo · in preparation for IJCAI 2027A move-by-move stopping rule and a symmetry-canonical search memory wrapped around KataGo. At the same mean visits Mikiri scores 78.8% over 1,000 games (+228 Elo), ahead of ten published stopping rules re-implemented on the same engine.Read PDF Code 
Audio Sliders: Measuring What a Slider Does to Musicsolo · in preparation for ISMIR 2027LoRA sliders for text-to-music diffusion on ACE-Step 1.5 and Stable Audio Open, each scored by measured audio descriptors and music-quality models, with newer axes found in 14,985 real recordings.Read PDF Code 
The Shrinking Edge: What Survives a Real Fill in a $97 Billion Prediction Marketsolo · in preparation for ACM EC 20271,009,373 resolved Polymarket markets and 23.7 million reconstructed fills, sent down a ladder of controls. Most claimed mispricing is measurement error; two effects survive, and one of them is fading.Read PDF Code 
Parallel Harmonic Exploration for Musical Accompaniment with MODULOwith P. Chin, E. Santos, N. SinghA within-subjects study of harmonic exploration and musical accompaniment in MODULO.Read PDF 
Long-Form Speech Reconstruction from Gradients in Federated ASRfirst author, with M. N. Bui and P. ChinAn experimental study of reconstructing speech features from gradients in federated ASR, using CTC loss, chunking, and error-driven optimization. Superseded by the closed-form recovery in Unsplice.Read PDF 
DataGo: Retrieval-Augmented Recursive Searchteam of sixA research prototype that adds selective retrieval and cached deep analyses to KataGo search. Superseded by Mikiri.Read PDF 
Unmixing the Music: Structured Recovery for Source SeparationsoloRobust PCA, NMF, K-SVD, and an unrolled deep sparse coder on one STFT front end, benchmarked against oracle masks on MUSDB18.Read PDF 
Operating-Point Control for Delegated Evidence Acquisitionwith A. MattarWhen should an agent escalate to a human? A prediction-market monitor tuned along its precision-recall operating frontier.Read PDF 
Belief-State Portfolio Allocation as a POMDPsoloA decision-theoretic trading overlay, reported honestly: the simple baseline wins on Sharpe, and the paper says so.Read PDF