Blog · 6 notes

Notes from the data layer.

Field notes, research and opinions from the 3Wheels AI team on codebase licensing, healthcare data, physical AI, speech and the craft of human data.

Codebase Data
Healthcare AI
Physical AI
Post-Training
Conversational AI
Data Annotation

Code editor open on a monitor in a dark room

Latest · Codebase Data · September 2026 · 5 min read

Why private production codebases are the next frontier for coding models

Read the note →

All notes

Every note, sorted by topic.

Codebase DataSeptember 20265 min read

Why private production codebases are the next frontier for coding models

Public GitHub has been scraped to exhaustion. The next leap in coding-model quality comes from licensed, production-grade private repositories, with tests, reviews and real bug history attached.

Read article →

Grid of CT brain scan slices on a lightbox

Healthcare AIAugust 20265 min read

Building healthcare datasets that survive regulatory scrutiny

De-identification is the floor, not the ceiling. What separates usable clinical data from a compliance liability is provenance: informed consent, patient-level linkage, and documented handling at every step.

Read article →

Robotic arms assembling a car chassis on a factory line

Physical AIAugust 20264 min read

Physical AI needs the real world, not just simulation

Sim-to-real transfer keeps hitting the same wall: the long tail of reality. Egocentric video, multi-sensor rigs, and real environments are how robotics foundation models close the gap.

Read article →

Server rack with patch cables lit green in a dark room

Post-TrainingJuly 20265 min read

Post-training is where models are won

Pre-training decides what a model knows; post-training decides what it is like to use. Preference data, expert demonstrations, and adversarial evaluation are the highest-leverage data you can buy.

Read article →

Broadcast microphone on a boom arm in a recording studio

Conversational AIJuly 20263 min read

Nine Indian languages, 37,000+ hours: inside our speech pipeline

Hindi, Bengali, Marathi, Malayalam, Tamil, Gujarati, Kannada, Telugu, and Indian English: real conversational audio at a scale that makes voice AI actually work for a billion speakers.

Read article →