Amin Abbasi

Ericsson · Ottawa

Built a graph neural surrogate that predicts RF cavity filter performance in under a second instead of five minutes of 3D electromagnetic simulation, plus a reinforcement learning agent that tunes filter geometry on its own. Taught myself the underlying physics from scratch to do it, and shipped the whole thing as a tracked, containerised pipeline the design team can run.

Nanda Lab, University of Ottawa · Ottawa

My PhD work. I led TEASMA and MetaSel — two frameworks for deciding whether a deep model's test set is good enough to trust, and for picking which unlabelled inputs are worth labelling. Currently using vision-language models to describe a model's blind spots in human concepts rather than as a number. Consistently beat the strongest published baselines.

University of Ottawa · Ottawa

Automated the mapping between software requirements and regulatory clauses. Compared fine-tuned classifiers, semantic similarity, and few-shot prompting — the prompting approach won using a fifth of the training data.

Giftpals · Cupertino, remote

First engineer on the product. Designed the architecture and database from nothing, then added the parts that made it interesting: a recommendation engine, customer segmentation, demand forecasting, and crawlers that gathered half a million product profiles.

Arnika · Tehran

Forecasting models for financial markets, sharpened with sentiment signals, and a reinforcement learning agent that traded on its own. Also built the live dashboards everyone actually looked at.

Shahid Beheshti University · Tehran

Where I fell for reinforcement learning: teaching a surgical robot to manipulate soft tissue, and combining evolutionary search with policy optimisation to make it converge. Two journal papers came out of it.