Skip to content
View takakhoo's full-sized avatar
7 followers · 15 following

Achievements

Achievement: Pull Sharkx3Achievement: Pair ExtraordinaireAchievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Pull Sharkx3Achievement: Pair ExtraordinaireAchievement: QuickdrawAchievement: YOLO

Highlights

  • Pro

Block or report takakhoo

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
takakhoo/README.md

Taka Khoo. Engineer, researcher, musician. Creative software, audio ML, and research systems.

Résumé (PDF) · Runnable experiments & results · Portfolio · Library · LinkedIn

AI and product engineer building creative tools, applied machine-learning systems, and production software. I have shipped music products used at scale, built research prototypes for DARPA INGOTS with NARF Industries and Dartmouth's LISP Lab, and completed theses spanning an AI-native DAW, neural audio restoration, and music composition and production.

My work centers on music, machine learning, signal processing, and product engineering. I care about systems that are useful beyond a demo: clear evaluation, honest limitations, human review where it matters, and documentation that lets another engineer reproduce the result.

Featured work

Each one has public code, saved results, and a manuscript you can read now (in preparation, not yet peer reviewed). Click a card for the paper.

Planning Through Regimes: a century of real-time evidence and a cube-root law for regime timing Planning Through Regimes · paper · code · explorer
Regime switching tested in strict real time since 1926: the published edges come from look-ahead, and the best recent jump model fails outside its sample. A POMDP with holdings in the state gives a cube-root no-trade band that matches dynamic programming within 2% and survives real trading costs.

Transposed Twins: benchmark leakage and memorization in symbolic melody models Transposed Twins · paper · code · listen
A transposition-proof twin search over 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 pays a 4-gram 0.55 nats per note against 0.01 to 0.06 for transformers, enough to reverse the ranking.

Lacquer: deciding what to fix in a finished mix Lacquer · paper · code · explorer
Measure first, then fix. Sparse declipping, cepstral echo removal (+16.4 dB), a band-split transformer for reverb (+5.9 dB where four released models gain at most 0.7 dB), and mastering held to the norms of 103,838 released tracks.

Unsplice: exact speech recovery from federated ASR updates Unsplice · paper · code · listen
One federated update from a speech recognizer gives back the client's audio in closed form: 98.5% of 1,417 LibriSpeech utterances, sequential decoding to 35 s, no transcript and no optimisation. Whisper reads the result at 3.8% WER.

Mikiri: knowing when a frozen Go engine has searched enough Mikiri · paper · code
A learned stopping rule and an exact search memory around a frozen KataGo. At the same mean visits it scores 78.8% over 1,000 games (+228 Elo), ahead of ten published stopping rules re-implemented on the same engine.

Audio Sliders: measuring what a slider does to music Audio Sliders · paper · code · live demo · weights
LoRA sliders on ACE-Step 1.5 and Stable Audio Open, each scored by measured audio descriptors and music-quality models. The newer axes were found in 14,985 real recordings rather than named in advance.

The Shrinking Edge: what survives a real fill The Shrinking Edge · paper · code
1,009,373 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 is fading.

MODULO studio session MODULO · modulomusic.com · thesis · user study
An AI-native music workstation heading to release: native Mac studio in C++ on JUCE and Tracktion, a SwiftUI iOS companion on TestFlight, and a FastAPI + Postgres backend on Render, Supabase, and Cloudflare with Stripe billing and Sign in with Apple.

Watch them run

Lacquer restoring a damaged mix, stage by stage Lacquer restoring a damaged mix: echo located and removed, reverb measured and left alone when it is music, gain ridden by a learned controller. Try the decision explorer.

Mikiri playing KataGo Mikiri as Black against KataGo: visits per move, memory hits, and its own estimate of the game.

Option model value against Polymarket fills The Shrinking Edge: a textbook digital-option formula on public spot data tracks Polymarket fills on a Bitcoin threshold contract.

Hear the sliders: Audio Sliders demo. Hear the attack: Unsplice players. Every demo is also playable in the interactive library on takakhoo.com.

Engineering focus

Python · C++ · Swift / SwiftUI · TypeScript · Objective-C++ · CMake · PyTorch · TensorFlow · React / Next.js · Node.js · FastAPI · PostgreSQL / Supabase / Neon · MongoDB · Stripe · Cloudflare · Docker · GCP · Render · Vercel · JUCE · Tracktion Engine · DSP · LaTeX

Recent work includes native Mac and iOS apps, real-time collaborative products, audio-model evaluation, LLM tool and agent systems, billing and provider backends, and graduate machine-learning instruction.

Elsewhere

Pinned Loading

  1. lacquer lacquer Public

    Automatic restoration and mastering for finished music mixes: DSP de-echo, band-split transformer dereverb, reverb decisions, learned level riding, mastering. Successor to my honors thesis.

    Python 1

  2. transformer-melody-generation transformer-melody-generation Public

    Readable TensorFlow encoder-decoder Transformer for symbolic melody generation.

    Python 3

  3. mikiri-beats-katago mikiri-beats-katago Public

    Mikiri: makes KataGo, the strongest open AlphaGo-style Go engine, 206 Elo stronger at the same search budget. No retraining.

    Python 1

  4. audio-diffusion-control audio-diffusion-control Public

    Slider controls for text-to-music diffusion: LoRA sliders on ACE-Step 1.5 and Stable Audio Open, scored by measured audio descriptors and music quality.

    Python 1

  5. prediction-market-research-agent prediction-market-research-agent Public

    Human-reviewed evidence discovery and monitoring system for prediction-market research.

    Python 1

  6. neural-audio-restoration neural-audio-restoration Public

    Token U-Net research for restoring degraded full-mix music in the neural-codec domain.

    Python 8 3