AutoMetal

AutoMetal is my Apple-M2-optimized Transformer inference implementation from TikTok TechJam 2026. I built it for a hackathon, then spent the rest of the project learning where the time goes: memory limits, shape-specific dispatch, fused attention, MPS and MLX behaviour, and correctness checks.

Across Shapes 1–13, the final implementation passed the challenge’s elementwise correctness gate with zero failed elements and reached a 4.98× geometric-mean speedup over the reference. I used Codex for the implementation, experiment design, benchmarking, and debugging. By the end, I understood the whole path from a kernel to a result I could trust.

AutoMetal overview diagram
Open AutoMetal · View the source on GitHub

Clinical Engine

Clinical Engine is an AI medical diagnosis simulator with virtual patients and physical exam modules. I built it so a learner can work through a case, ask questions, and decide what to do next instead of reading the answer first.

Clinical Engine
Watch the demo video.

XAIR

XAIR is an explainable-AI pipeline for multi-domain reasoning. It generates multiple reasoning paths, finds where they diverge, tests counterfactual changes, and checks claims against Wikidata.

I worked on it during my undergraduate research at NTU because I wanted to understand what a model’s reasoning trace can tell us, and where it starts to mislead us.

The Handbook for the Enterprising Teen

I wrote The Handbook for the Enterprising Teen for teenagers with an idea and no clear first step. It is a practical guide to starting small, testing ideas, and learning by doing.

The Handbook for the Enterprising Teen book cover
Find it on Amazon.

The Veer Dosi Show

The Veer Dosi Show is my podcast about entrepreneurship and personal development. I use conversations to ask how people started, what they got wrong, and what they do when the first version of the plan stops working.

The Veer Dosi Show podcast artwork
Listen to the show.