Daniel Goldberg
Software · Machine learning · South Florida

I build things that ship — with AI in the loop.

AI that designs DNA edits. A machine that hears piano. Platforms running in production while I sleep. I don't stop at the prototype — I run the infrastructure, the automation, and the 2 a.m. deploys.

Scroll
01

Research & builds

The deep stuff — models, sensing, and infrastructure built from first principles. Most of this runs locally; public repos are linked where they exist.

01 · Genomics · Science-fair research

AI-assisted design of targeted mutations

A model that reads DNA and writes the smallest edit that produces a chosen effect — not just predicting what mutations do, but designing the ones you want. Scored 92% exact-match — seven points past the original 85% goal.

Science fair · full paper written Data & notebooks
What it is

An encoder–decoder Transformer that takes a DNA sequence plus a natural-language target effect and generates the minimal base-pair changes, in HGVS form. Trained on ~260,000 real human variant sequences at 2,056 bp each, built from human-reviewed UniProt annotations, Ensembl cDNA, and Levenshtein-inferred codon changes.

What I did

The whole pipeline: multi-threaded UniProt/Ensembl scrapers with graceful HTTP retries, a custom-trained parser that turns free-text variant effects into training conditions, K-mer + SentencePiece tokenization, and an architecture that makes 2K-bp DNA tractable — RoPE, grouped-query attention, and an 8× CNN downsample. Fifteen full training runs on TPU v5e; the dataset later grew to 840,000 samples.

92.19%avg test accuracy
92.15%exact-match accuracy
<0.9%variance, 15 runs
TensorFlow · PyTorch · Transformers · RoPE · GQA · UniProt / Ensembl APIs · SentencePiece
Data & notebooks
02 · Audio ML · Piano transcription

better-midi — teaching a machine to hear piano

Solo piano audio goes in; the MIDI comes back out. A Conformer transcription model built and trained from scratch on MAESTRO — no pretrained backbone, no borrowed weights.

clean eval run in progress Kaggle
What it is

An audio→MIDI piano transcriber: a 6-layer Conformer (d_model 512, macaron feed-forwards, 31-tap convolution module) over mel spectrograms, with onset, frame, and velocity heads that decode back into 88-key piano rolls and note events.

What I did

Built the RAM-cached chunking pipeline (79,563 training chunks from MAESTRO v300 — 300 virtuoso performances), the training loop with SpecAugment-style masking and cosine LR decay, and a 50 ms-tolerance note evaluator on mir_eval. The model trains end-to-end from scratch.

precision @ 50ms
note recall @ 50ms
F1 score @ 50ms
PyTorch · Conformer · torchaudio · MAESTRO · mir_eval · note-seq
Kaggle
02

In progress

In progress · IP·01

AI wristband

A wrist-worn AI device — hardware, firmware, and on-device machine learning, all in active development.

Telemetry

Hardware, firmware, and on-device machine learning — bring-up happens as you scroll: sensors online, model streamed to the edge, telemetry live. No screen on the band — every readout here lives in the scene.

Edge ML · Firmware · On-device inference
Building now
In progress · IP·02

Multi-instrument MIDI transcription

better-midi's big sibling: transcribing whole ensembles, not just solo piano — instrument-conditioned, autoregressive decoding across tracks (drums included) against multi-instrument ground truth.

The hard part

Whole ensembles, not solo piano — instrument-conditioned, autoregressive decoding across tracks, checked against multi-instrument ground truth. Watch the ensemble become a score as you scroll.

PyTorch · Autoregressive decoding · Multi-track MIDI
Building now
03

More builds

More builds · M·01

Garden app

Writeup in progress.

Details soon Personal project
Night garden

A rotating specimen ring under a moonlit sky — scroll to turn the carousel, click any plant to bring it to the front. Full writeup coming soon.

Personal project
Details soon