Hi! I am Usman! I am currently fucking around and finding out at SF Parc. For the past few months I was building conviction and trying to build a startup around AI and Bio at Entrepreneurs First.

Some things I am interested in currently: risk management for ai, smol LLMs, and AI for scientific discovery. I am especially interested in building new software and hardware to reimagine the scientific instrument stack, rather than just wrapping LLMs around existing workflows or doing end-to-end ai agents. Matlab/MathWorks pulling in $1B+ a year for software everyone complains about shows how badly science and engineering needs better tools.

Prior to taking a bet on myself, I was working as a lead quantitative/ML researcher at Bilby AI where I was building ML models to understand and predict policy changes and the downstream impacts on financial markets, providing actionable insights for G7 nations, petrostates, and some big hedge funds. My research stack at Bilby covered everything from classical generalized linear models to fine-tuning LLMs. Before working at Bilby, I used to be a quantitative researcher at TAS Alpha where I worked on building models for binary options, order-book modelling, distributionally robust portfolio optimization, and model calibration.

Outside work I spend my time reading weird ML Papers with a focus on NTKs, grokking, and SAEs for interpretability. Whatever time I have left goes to reading, filling up my journal with mundane thoughts, learning how to make a good ribeye, and taking long walks.

Past Life

I graduated from City University of Hong Kong in 2022 where I studied mathematics.

In the past, before graduation I tinkered with different spaces and played around a lot. In the industry I wore the hat of machine learning researcher and engineer, where I worked on building computer vision models for classification and object detection (the good old days of EfficientNet, when VITs were still new) on researching and deploying machine learning models at Dayta AI. I also worked on building graph and ML systems for programmatic advertising at AlikeAudience. On the engineering/building side, I built ugly beautiful frontends (I was subsumed by React), created mess on the backend (for a minute I loved GraphQL too), deployed insecure apps (banged my head against Apache, then later realized that nginx was a better choice), and then settled for building, deploying, and researching machine learning models à la implementing and simplifying a lot of ML papers and making use of all the distributed tricks up my sleeves to speed the GPUs.

However, my heart mostly belonged to mathematics. I spent a bunch of my time thinking about problems in quantum information where my research interests in general spanned quantum complexity theory and quantum shannon theory. During my last project in academia I worked with Professor Marco Tomamichel at NUS where my focus was on using tools from real semi-algebraic geometry for quantum resource theories.

If you are interested in this area for starters I recommend looking at the following papers/surveys:

• Asymptotic spectra: Theory, applications and extensions
• Quantum Resource Theories
• Resource convertibility and ordered commutative monoids

Although the mysteries of my life led me to leave academia, here is a small list of the publications and talks I worked on:

• (Publication) Matrix Majorization in Large Samples in IEEE Transactions on Information Theory
• (Publication) Generic detection-based error mitigation using quantum autoencoders. Physical Review A, 103(4).
• (Talk) Quantum Communication Complexity and Quantum Information Complexity @ NYC Quantum Meetup
• (Talk) Vergleichsstellensätze and Quantum Resource Theories