This is the website of Alif Munim. I am a machine learning specialist at the University Health Network, building AI foundation models for cardiology & clinical trial matching. I am best known for my work on EchoJEPA; MedRAX; and AI safety & alignment.
Previously, I worked on IBM Watson; interned at the Vector Institute; and conducted research at the University of Toronto & Toronto Metropolitan University. I graduated with distinction with a bachelor of science (BSc) degree in computer science in 2023, with a minor in psychology. I also serve as Community Lead for AI Safety & Alignment at Cohere Labs.
| 2026 | EchoJEPA: A Latent Predictive Foundation Model for Echocardiography arXiv preprint arXiv:2602.02603 · arXiv · website |
| 2025 | MedRAX: Medical Reasoning Agent for Chest X-ray arXiv preprint arXiv:2502.02673 · arXiv |
| 2025 | Benchmarking and Adapting On-Device Large Language Models for Clinical Decision Support arXiv preprint arXiv:2601.03266 · arXiv |
| 2023 | Interpretable Machine Learning for Automated Left Ventricular Scar Quantification in Hypertrophic Cardiomyopathy Patients PLOS Digital Health 2(1), e0000159 |
| 2023 | Bringing the State-of-the-Art to Customers: A Neural Agent Assistant Framework for Customer Service Support EMNLP 2022 (Industry Track) |
Built on Yann LeCun's JEPA architecture, EchoJEPA predicts the meaning of masked cardiac video regions in an abstract latent space rather than reconstructing pixels. Trained on 18 million ultrasound videos from 300K patients at UHN. 20% lower error on cardiac function estimation; 79% view classification with 1% of labels; zero-shot pediatric transfer beats all baselines.
The first AI agent framework that integrates specialized chest X-ray analysis tools with large language models via a ReAct reasoning loop. Coordinates CheXagent, MedSAM, and other clinical tools to perform complex multi-step medical reasoning. Achieves SOTA on ChestAgentBench (63.1%) and SLAKE VQA (90.35%) without additional training.
Evaluating whether on-device LLMs (20B and 120B parameters) can match cloud-based systems for clinical tasks while preserving patient privacy. Designed three evaluation frameworks: generalist medical reasoning, specialist clinical scoring, and LLM-as-judge. Fine-tuning the 20B model boosts accuracy from 77% to 87%, approaching GPT-5 and beating DeepSeek-R1.
An investigation into diffusion models for generating realistic medical images to address data limitations in healthcare. Experiments showed that training from scratch on domain-specific data consistently outperformed fine-tuning pretrained natural-image models. Reviews GANs, VAEs, and diffusion model foundations.
Compared U-Nets, fully convolutional networks, and gradient-boosted ensemble models for brain metastasis segmentation in <100 3D MRI scans. Deep learning significantly outperformed classical methods; data augmentation proved critical for generalization.
Extended speech-act theory into deep learning. Built taxonomic word embeddings for email intent classification using Carvalho & Cohen (2005). This work contributed to IBM Watson Orchestrate, which won a CES 2022 Innovation Award.