# Dilawar Mahmood

> Dilawar Mahmood is a machine learning engineer at ZeroEntropy (YC W25) in San Francisco, best known for four years at Apple where he optimized on-device models for Siri and Spotlight - work he once presented directly to Tim Cook at the Steve Jobs Theater. A Norwegian-educated engineer who left a comfortable career track to attend the Recurse Center and rediscover what programming actually feels like, he builds distributed ML frameworks in his spare time and is on record hating vibe coding.

- **Role:** Machine Learning Engineer at ZeroEntropy (YC W25)
- **Organizations:** ZeroEntropy, Apple, Recurse Center
- **Nationality:** Norwegian
- **Education:** Bachelor's in Computer Science, Norwegian University of Science and Technology (NTNU)
- **Known for:** Presented on-device AI model improvements to Apple CEO Tim Cook and AI/ML leadership at the Steve Jobs Theater, Led multilingual and on-device transformer optimization work for Siri and Spotlight at Apple for 4 years, Built distributed-hetero-ml: open-source framework enabling model training across heterogeneous hardware (Apple Silicon + NVIDIA GPUs)

## Career timeline

- **2019-2021** — Student hire at a tech company while studying at NTNU; bachelor thesis on federated learning with differential privacy and homomorphic encryption
- **2021-2025** — Lead Machine Learning Engineer at Apple, Barcelona - worked on Siri, Spotlight search, foundation models, on-device transformer optimization, and multilingual capabilities
- **2023** — Presented on-device model performance improvements to Tim Cook and Apple AI/ML leadership at Steve Jobs Theater, Cupertino
- **2025** — Left Apple to attend Recurse Center (S2'25 batch) in Brooklyn, NYC - worked on distributed training across heterogeneous hardware and pre-training hybrid U-Net Transformer architectures
- **2025** — Joined ZeroEntropy (YC W25) as Machine Learning Engineer in San Francisco

## Achievements

- Presented on-device AI model improvements to Apple CEO Tim Cook and AI/ML leadership at the Steve Jobs Theater
- Led multilingual and on-device transformer optimization work for Siri and Spotlight at Apple for 4 years
- Built distributed-hetero-ml: open-source framework enabling model training across heterogeneous hardware (Apple Silicon + NVIDIA GPUs)
- Pre-trained a hybrid U-Net Transformer architecture under GPU-constrained conditions at Recurse Center
- Bachelor thesis on federated learning incorporating differential privacy and homomorphic encryption
- Implemented AlphaZero algorithm for board games (Tic-Tac-Toe, Four in a Row)
- Arctic Code Vault Contributor (GitHub)

## Latest updates

- **2025-10** — Completed Recurse Center S2'25 batch in Brooklyn after 3 months focused on distributed training and transformer pre-training research
- **2025-10** — Published 'Pre-Training a Hybrid U-Net Transformer' and 'Recurse Center Return Statement' on dilawar.ai
- **2025-07** — Published 'Being GPU Poor makes you creative' - open-sourced distributed-hetero-ml framework
- **2025-06** — Published 'Vibe Coding' post arguing for AI as copilot, not pilot, with co-author Michi (his cat)
- **2025-01** — Listed as team member at ZeroEntropy (YC W25), the YC-backed AI retrieval infrastructure company

## Links

- Website: https://www.dilawar.ai
- LinkedIn: https://www.linkedin.com/in/dilawar/
- GitHub: https://github.com/dilawarm

---

Profile page: https://yespress.io/dilawarmahmood
Published by YesPress — https://yespress.io
Last updated: 2026-04-17