Machine Learning Engineer - ZeroEntropy (YC W25)
"I'd rather understand my code than vibe with it."
From NTNU lecture halls in Norway to the Steve Jobs Theater in Cupertino - and now building AI retrieval infrastructure at a Y Combinator-backed startup in San Francisco. He is the engineer who left Apple on purpose, and whose cat Michi has a blog credit.
Profile
Most people do not voluntarily leave a Lead Machine Learning Engineer role at Apple. Dilawar Mahmood did. Not because he was fired. Not because he had a better offer. Because he felt his skills eroding under a slow accumulation of alignment meetings, stakeholder syncs, and coordination overhead. He left because he missed the actual thing - the hard, dirty, satisfying work of building something from scratch and understanding every layer of it.
That decision says more about him than his four years of Apple credentials ever could.
"I felt like I was regressing as a programmer."- Dilawar Mahmood, on leaving Apple
Before the career pivot, though, there was the career itself. At Apple's Barcelona office, Mahmood spent four years doing serious work - on-device transformer optimization for Siri, batch inference infrastructure, multilingual foundation models for Spotlight search. This was not peripheral AI work. This was the stack that runs in hundreds of millions of pockets. And at some point he stood in the Steve Jobs Theater in Cupertino and presented performance improvements for on-device models to Tim Cook and Apple's AI leadership. He described it afterward as "terrifying and surreal." Two accurate words.
The foundation was built in Norway. He studied at the Norwegian University of Science and Technology (NTNU) in Trondheim, where his bachelor thesis combined federated learning with differential privacy and homomorphic encryption - the kind of project that signals either an overachiever or someone who finds genuinely hard problems more comfortable than easy ones. Probably both. He worked as a student hire for two years during his studies, demonstrating early that he could operate independently on challenging assignments without needing much supervision.
After Apple came the Recurse Center in Brooklyn - a three-month programmer's retreat in New York City where participants work on whatever interests them, surrounded by other people who care about programming as a craft. For Mahmood, this was the recalibration. He explored distributed training across heterogeneous hardware (Apple Silicon M4 alongside NVIDIA TITAN X cards - a setup that should not work as a unified training cluster and mostly does not, until you apply Ray and some creative engineering). He pre-trained a hybrid U-Net Transformer architecture from scratch under real GPU scarcity. He played chess and Love Letter and Decrypto with fellow participants, and pioneered something he called Film Night as a lower-stakes alternative to serious cinematic discussion.
He also published a blog post about vibe coding that landed with some force among people who think about how engineers interact with AI. His position: using AI to generate code you do not understand is not engineering - it is gambling with debt payments due at debugging time. The essay is precise, measured, and written with a cat. His cat Michi appears in the byline as co-author.
"Ask the AI and hope. That's not engineering, that's gambling."- Dilawar Mahmood, "Vibe Coding" (2025)
The current chapter: ZeroEntropy, a Y Combinator Winter 2025 startup building retrieval infrastructure for AI systems. Embedding models, rerankers, managed search - the plumbing that determines whether AI products can actually find the right context at the right time. Mahmood joined as a Machine Learning Engineer, bringing his on-device optimization experience and a hard-won respect for systems that behave predictably under pressure.
He maintains a personal site at dilawar.ai where his writing tends toward the technically dense - blog posts on multi-cluster distributed training across heterogeneous hardware, U-Net Transformer pre-training on a shoestring GPU budget, and the epistemics of what it means to understand the code you deploy. The site has one more page than you expect, and the author bio is two words: "Hi, I'm Dilawar."
On GitHub as dilawarm with 47 repositories and an Arctic Code Vault Contributor badge from a time when GitHub literally froze code in Arctic permafrost. He builds frameworks that abstract away complexity, implements AlphaZero for board games because why not, and releases things with the quiet confidence of someone who just wants useful things to exist in the world.
The through-line across all of it is a stubborn commitment to understanding. Not fluency - understanding. The kind that comes from sitting with a system long enough to know why it fails, not just how to patch it. In an era of AI-generated everything, that orientation is either charmingly anachronistic or quietly essential. Probably the latter. People who understand their tools deeply tend to build things that last.
Career
Achievements
Presented on-device Siri performance improvements to Apple CEO Tim Cook and AI/ML leadership at the Steve Jobs Theater - "terrifying and surreal" in his own words.
Pre-trained a hybrid U-Net Transformer architecture from scratch under real GPU-constrained conditions at the Recurse Center - published full writeup on dilawar.ai.
Built distributed-hetero-ml: an open-source framework that enables model training across incompatible hardware (Apple Silicon + NVIDIA) simultaneously via Ray.
Bachelor thesis at NTNU implemented a federated learning framework incorporating both differential privacy and homomorphic encryption - an unusually thorough combination for undergraduate work.
4 years as Lead ML Engineer at Apple working on foundation models, Siri, Spotlight search, and multilinguality - on-device AI at global scale.
GitHub Arctic Code Vault Contributor - his code is preserved in Arctic permafrost for future civilizations. Michi's co-authorship credit is, unfortunately, not included.
In His Own Words
I'd rather understand my code than vibe with it.
There's a big difference between using AI as an assistant and letting it drive everything.
If you outsource all that thinking to AI, you're not developing those skills.
Good software comes from understanding what you're building and why.
It's about understanding how something works, solving hard problems, and building things with people you respect.
Use AI as a copilot, not the pilot. Let it handle tedious stuff while you focus on system design, architecture.
Extras
His GitHub username is dilawarm and his LinkedIn handle is just "dilawar" - the man respects brevity in identifiers.
His cat Michi is credited as co-author on his "Vibe Coding" blog post. Whether Michi contributed the anti-AI-slop argument or the cat photos is unclear.
Norwegian by education, Spanish by four-year tenure (Barcelona), American by current zip code (San Francisco). Three countries, one passport.
Between Recurse Center coding sessions he played chess, Love Letter, and Decrypto - which is either a sign of excellent taste or a coping strategy.
His bachelor thesis crammed in both differential privacy AND homomorphic encryption. Most students pick one. He did both, presumably for fun.
Invented Film Night at Recurse Center as a deliberately low-pressure alternative to serious movie discussions. Community building through deliberate casualness.
Connect