Almutwakel Hassan
I'm building 3D video generation models at True3D.
Previously, I studied Artificial Intelligence at Carnegie Mellon University and interned at Meta and Citadel Securities.
Timeline
Feed-Forward 4D Video Generation
At True3D, I am designing temporally consistent feed-forward generative models for 4D Gaussian splat sequences. This direction was motivated by fundamental limitations I discovered in per-scene optimization approaches -- physically-motivated motion representations underperform simpler learned functions for 4D Gaussians, because reconstruction is too locally optimized to recover real motion without strong external priors like 4D point tracking. We are utilizing this to create 3D representations of movies watchable in VR.
4D Gaussian Splatting Reconstructions
I built 4D Gaussian Splatting reconstructions at True3D, finding that learned time functions with volume-maximizing regularization outperformed per-frame optimization at 100x less storage. Without this regularization, Gaussians converge to degenerate solutions -- shrinking and darkening when optimized across time. I ran systematic ablations across conditioning strategies, VAE architectures, and attention patterns to produce temporally consistent volumetric sequences. We created high quality streamable 4D Gaussian splats playable within the web browser.
Scaling 4D Diffusion Training
I built and iteratively refined a generative model architecture from scratch at True3D, training 1B+ parameter flow matching transformers on high-resolution 4D data. I own the entire pipeline -- sourcing and pre-processing large video and 3D datasets, distributed training across 24 H200 GPUs on AWS and Google Cloud, automated experiment scheduling, cluster orchestration, and deploying the model behind an inference endpoint. The role demands ownership of the entire research to production pipeline, from business-constrained planning and background research to containerized deployment. We created a model that generated sparse 4D voxel videos given a text prompt (e.g. "A dog running and barking").
Auto-Bidding Algorithms for Ads
At Meta, I worked on the automatic bidding algorithm that handles 95% of revenue in their ads auction marketplace. I proposed and implemented a solution for fast convergence to equilibrium bid price during the cold start period, and engineered a machine learning model ETL pipeline using campaign data to predict optimal bidding price targets. The algorithm changes produced massive revenue impact, validated through A/B tests and rolled out globally across the platform.
Self-Repairing Robot Manipulation
For my senior research thesis at the Bot Intelligence Group, I focused on using large language models and vision-language models to repair robot manipulation task execution. I built task monitoring agents that evaluate whether executions failed and determine the source of failure. My approach has the agents analyze the environment visually to determine the expected intermittent and final states, given the context of the task and robot, and assess whether these states are achieved.
CMU Poker AI Competition
As president of the Data Science Club, I envisioned hosting CMU's first ever poker bot competition and worked tirelessly to make it a reality. The technical challenges were immense -- players submit Python code to GitHub for their bots, our systems automatically build it into a Docker container with any dependencies and send it to our bots database. We then have a constant stream of matchmaking that deploys these containers and runs them against each other in the cloud, calculating match results, player statistics, and ELO scores, and forwarding the results to the web servers.
My goal was to enable as many different types of approaches as possible, including algorithmic, mathematical, rule-based, and offline and online machine learning approaches. I led the technical efforts along with various area leads to architect the system, personally contributing thousands of lines of code. The end product was over 30,000 lines of code in Python and Typescript. I led a team of 20 developers and 12 organizers to bring the event to fruition for over 400 participants, and I secured $100,000 in prizes for future iterations.
AI News Analysis
I achieved first place among 100+ teams in the Rutgers RAISE 2024 Data Science Competition by analyzing articles to assess media portrayal of AI. I used Sentence Transformers, NLTK, and the OpenAI API to conduct sentiment analysis, clustering, and LLM-based categorization.
Solar Energy Forecasting
I competed in the DOXA AI ClimateHack, forecasting solar energy generation using Transformers and CNNs with time series satellite and weather data. I presented at Harvard University representing CMU and achieved 6th place among all schools in the US, UK, and Canada.
Data Science Club at CMU
I led the Data Science Club at Carnegie Mellon, an organization dedicated to providing educational experiences in data science and artificial intelligence. When I took over as president the club had about 300 members. I organized and personally taught several educational workshops, created a mentor-mentee matching program, and ran semesterly competitions for hundreds of students. I personally taught events like our deep learning workshop, mentored 25 students in weekly hands-on Kaggle challenges across three semesters, and directed teams of organizers and developers to successfully launch our competitions for the first time. By the time I stepped down, we had over 1,000 members -- making it one of the most wide-reaching student organizations at CMU. I am truly invested in the data science and machine learning education of my peers and I work hard to bring my visions to life.
Trading Data Infrastructure
At Citadel Securities, I created a versatile trading data retrieval API enabling aggregate analysis across millions of profit/loss events to support teams globally. I designed a highly optimized and robust querying system to significantly accelerate real-time data retrieval. The internship gave me exposure to the quantitative finance industry and strengthened my data engineering and systems skills.
Cross-Embodiment Robot Manipulation
I joined the Bot Intelligence Group, a robotics research lab at CMU, in my first year. Under the supervision of Tanmay Shankar and Professor Jean Oh, I dedicated my first research project to simulating, training, and analyzing robot manipulation for highly dexterous robot hands trained from human demonstrations using unsupervised machine translation methods. I demonstrated a way to translate manipulation skills from human hand movements to robot hand movements without explicit mappings between them through the use of unsupervised autoencoding methods and generative models. The work involved developing data processing pipelines, creating simulators for visualizing actions, running training experiments on both action space generation and translation models, and evaluating the framework in each domain.
I continued to work on incorporating more environmental information into skill representation and exploring reinforcement learning methods within the framework. I led ablation studies isolating which components of the representation drive sim-to-real performance. This work was published as "Translating Agent-Environment Interactions across Humans and Robots" at IROS 2024.
Autonomous Racing Perception
I developed perception systems for CMU's Carnegie Autonomous Racing team's self-driving race car. One particularly challenging problem I worked on was to reduce the density of LiDAR point clouds for increased processing speed without critically reducing the accuracy of clustering algorithms for obstacle detection. I also worked on optimizing YOLOv5 camera pipelines for the vehicle.
Opioid Withdrawal Prediction
My first internship was at Behaivior, a Pittsburgh startup focused on predicting opioid cravings using biometric data to deploy preventative measures for patients with opioid use disorder. I focused on dimensionality reduction, feature selection, and model architecture design for the company's most important models. I improved prediction accuracy of time series biometric data CNN models by 9% and integrated the data pipeline and model inference with Google Cloud services to deploy models to the app backend. My models predicted when cravings were likely to occur using live biometric data, so that we could deploy preventative measures in time to help people before relapse can occur. This work was very meaningful to me, and I am proud to have had a long lasting effect on the company as they work towards their mission.
Hackathons and Game Dev
I won HackCMU for Personalate (More Connected World category) and TartanHacks 2022 for Pnumber (Unforgettable Hack category). I also won the Game Creation Society 2021 Fall Games Showcase for Lingua Litis, a game I built with a group of developers, artists, and musicians using Unity. I've been motivated to create my own projects ever since my junior year of high school, and I've created some in several different domains to explore the capabilities of the field.
Psychology Behind Charity
I began a research project in my first semester at CMU, applying machine learning and data science methods to a psychology project. This project explored links between the language usage of charitable foundations and the resulting charitable giving. Using large datasets, web-scraping, language analysis tools, Python and R, I created predictive models and data summaries to determine the effects of natural language on charity outcomes.
US National AI Commission Testimony
I presented testimony sharing my views on Artificial Intelligence societal impact and regulation to Congressional Representatives Mike Ferguson and John Delaney, along with AI experts, at the National AI Commission hosted by the US Chamber of Commerce. I was quoted twice in a published article from the US Chamber of Commerce. One of my lifetime ambitions is to help ensure the inevitable mass usage of AI in automation is used to benefit people in need, especially people from underrepresented communities.
Carnegie Mellon University
B.S. in Artificial Intelligence from the School of Computer Science, 3.88 GPA. I have been building my experience in machine learning since high school through self-directed projects. My coursework at CMU prepared me across deep learning, reinforcement learning, computer vision, generative AI, distributed systems, cloud computing, and robot planning.
Botball National Champions
I led my team as captain to win first place nationally in KIPR's autonomous robotics Botball Competition, leading strategic design, programming, and engineering efforts towards 2 autonomous robots.
Secaucus High School
Graduated as valedictorian with a 4.73 GPA. I founded the Chess Club, served as president of Mu Alpha Theta (Math Honors Society), and captained the Robotics Team. I spent a lot of my time tutoring other students in math and computer science. I've been motivated to create my own machine learning projects ever since my junior year of high school, and I've created some in several different domains to explore the capabilities of the field.