Founder of Clef, a piano learning app that schedules practice the way Anki schedules flashcards. It builds on my research in AI systems, optical music recognition, and tools that accelerate human learning.
A web app for memorizing piano repertoire. You play over MIDI and every note is graded. FSRS spaced repetition decides which bars to review, the same algorithm Anki uses for flashcards.
In production with the full loop live. MIDI grading in the browser, FSRS scheduling, score catalog, and billing. The catalog is populated from public-domain scans by Transcoda, the OMR model below.
An optical music recognition system built around synthetic data generation, normalized Humdrum **kern output, and grammar-based decoding for syntactically valid transcriptions.
A compact 59M-parameter model trained in six hours on a single GPU, outperforming billion-parameter baselines on synthetic and historical scan benchmarks.
End-to-end design and delivery: catalog ingestion, a custom keyboard-first labeling tool, active learning with a Cleanlab label audit, and a category-conditioned DINOv3 ViT-L classifier (304M parameters) served in production over gRPC.
26,920 human labels → macro-F1 0.82 across 11 photo-type classes; the full 1.65M-image catalog scored in 11.5 hours on a single GPU, 88% of predictions above 0.9 confidence.
A C++20 implementation of the OMR-NED score-difference metric for evaluating optical music recognition systems — Verovio-based parsing of MusicXML, MEI, and Humdrum/kern, a musicdiff-compatible edit-distance engine, and annotated side-by-side visual diff reports.
A personal observability platform in Rust — eight capture channels (Anki, e-reader, Fitbit, MIDI, mpv, browser) feed one HTTP API into SQLite, coming back out as goal-scored plots, streaks, and EWMA-based trend analysis.
A high-performance and low-memory popup dictionary application in C with GTK3 that operates on text selection and automatically generates Anki flashcards.
High-performance Japanese word deinflection using perfect hash tables and constant-time lookups.
A command-line workflow for creating Anki cards from PDF reading sessions with contextual source data.
