hi, I'm Muhiim
It means "important" in Somali. I'm working on living up to that.
I'm a recent Brown CS grad with experience across machine learning, computer vision, and NLP, and I've spent the last few years doing AI research and teaching computer science at Brown. I'm passionate about how humans learn, how machines learn, and how we can build technology that helps people access knowledge and opportunity.
Outside of work, I enjoy calisthenics, soccer, strength training, and chasing a runner's high.
Work
Research
DRIVE: Dynamic Relation Inference via Verb Embeddings
A new method for improving how vision-language models understand relationships between objects in images. Built CROCO, a refined dataset from COCO that distinguishes static from dynamic visual relationships, and designed a contrastive loss function that outperforms existing approaches on the task.
CLEAR GLASS: Contrastive Learning with Enhanced Abstract Representations
A method for teaching vision-language models to recognize abstract concepts beyond objects and relationships. Introduces MAGIC, a grouped image-caption dataset, and a novel grouped contrastive loss that induces models to encode higher-level semantic abstractions as an emergent capability — without ever being shown the higher-level concepts directly.
Projects
Does Curiosity Shape Transferable Representations?
Systematic study of intrinsic motivation methods — ICM, RND, CFN, and NLL-Surprise — as representation learning mechanisms for cross-task transfer in Procgen RL environments. Analyzed 507 checkpoints across 12 games to evaluate how curiosity-driven exploration affects the generalizability of learned features. Supervised by Professor George Konidaris.
Avatar: 3D Human Digitization Pipeline
End-to-end pipeline for turning real people into animatable 3D avatars. Captured multi-angle depth data with an Azure Kinect, fused point clouds in CloudCompare via ICP, reconstructed surfaces using the Ball-Pivoting Algorithm, refined meshes in Blender, and fitted SMPL-X with kNN skinning weight transfer to drive motion-capture animations.
Cryptographic Voting for Multiple Candidates
Secure voting system for k-of-t elections using additively homomorphic ElGamal encryption, zero-knowledge proofs, blind RSA signatures, and threshold decryption. End-to-end verifiable while preserving ballot secrecy.
BERT Question-Answering System
Fine-tuned BERT for extractive QA on the Natural Questions benchmark. Custom loss functions and preprocessing brought competitive F1 (53.8%) without architectural changes.
Reddit Depression Detection
ML pipeline to detect depression signals in Reddit posts, combining LDA topic modeling and RoBERTa embeddings as features for random forest classifiers.
Craigslist Price Predictor
Multimodal price prediction combining VGG16 image features and spaCy text embeddings with gradient boosting. Handles both visual and textual listing data.