Nilesh Patil

AI systems & applied research. I’m Head of AI at DreamStreet, building compliance-aware AI for SEBI-regulated (India’s securities regulator) investor and trader workflows. Before that I led applied AI research at Dream11 and ran a Columbia University research collaboration.

This site is where I write up the engineering behind that work - mostly field reports on HPC, applied ML, and the measurements behind the results.

About →  ·  Publications →  ·  Side projects →  ·  CV →  ·  Search →

Latest · 16 Apr 2026 · 20 min · about 2x the cells per compute budget Adaptive OpenMP scheduling for tissue-mechanics simulations A tissue simulator’s threads sat pegged at 800% while the trace put roughly a third of that time idle at a barrier. A fork that rebalanced the OpenMP schedules and rebuilt the contact grid reached about 2x the cells per compute budget, and the runs cannot say how much of that belongs to the scheduler. cpphpcopenmp

All 14 posts →  ·  Posts by year →  ·  Posts by tag →

Side projects

SimuCell3D: HPC tissue mechanics in C++ Performance-oriented fork of an ETH-developed C++ framework for 3D tissue mechanics at subcellular resolution. In two benchmark runs it ran 1.5x to 2.75x faster than v1.0 at matched cell counts and reached about 2x the cells in the same wall-clock time. GitHub ↗
Ultrasound Doppler angle from B-mode with deep learning A deep-learning project that reads the Doppler angle from a single grayscale carotid B-mode image. It extends my EMBC 2019 paper with grid pooling on a frozen backbone, Optuna-tuned estimators under two sampling protocols (down to ~2° MAE), and a calibrated clinical evaluation. GitHub ↗

All side projects →