I'm an AI research engineer in Islamabad, Pakistan. Most recently I was at Unify (YC W23), where I built much of the platform under its AI teammates, from the RAG pipeline to the generative UI they acted through and their integrations with a company's apps, and then led a ground-up redesign of its open-source agent harness, with an agent mesh and a memory built for both memoization and generalisation. The new harness solved more Continual-ARC puzzles than the old one and than OpenClaw, Hermes Agent, and Prime Agent, though, like them, it got no better at puzzles it had never seen, and I'm a co-author of the benchmark's paper. More on that work.
Before that I co-founded Ivy and led its source-to-source ML transpiler, which rewrote PyTorch models and libraries as working, readable TensorFlow and JAX, so a PyTorch model could ship to phones through TensorFlow Lite as source an engineer could still edit; its integration was merged into Kornia. I started out at CureMD, training layout-detection and entity-recognition models for document understanding.
I'm looking for my next role: building agents and their harnesses and running them securely and reliably, or post-training the models inside them, which is where I've been spending my time since. What draws me to it.
- Unify agent harness: the open-source harness in which Unify's agents act by writing and running Python. I led its ground-up redesign, with an agent mesh and a memory built for both memoization and generalisation.
- Ivy: the transpiler I led, from the traced-graph version at Unify to the source-to-source rebuild at Ivy, which rewrote PyTorch code as readable TensorFlow and JAX with its control flow intact.
- Kornia: co-authored the integration, merged in October 2024, that gave this PyTorch computer-vision library TensorFlow, JAX, and NumPy support through Ivy's transpiler.
- TensorLy: co-mentored the Google Summer of Code 2023 project that brought TensorLy-style tensor decompositions into Ivy.