I have a passion for deploying AI, ML, and Optimization models in critical industries, at scale. Over the past few years, I solved problems in cybersecurity, transportation and infrastructure. I expanded my background from Operations Research to LLMs, Multimodal AI, and Reinforcement Learning.
Almanax: AI for Cybersecurity
For the past ~1.5 years, I have been working on Almanax. I was the first hire and my main responsibility has been to lead AI research and engineering. Almanax is a platform that allows enterprise security teams to deploy AI Agents for vulnerability detection, triaging and remediation. Our AI Agents achieved SOTA performance on security engineering tasks and are loved by hundreds of teams worldwide. We also pioneered high quality benchmarking for cybersecurity.
Computer Vision Research
Let's start from one of my biggest commitment so far, my research work at MIT. Under the guidance of prof. Alexandre Jacquillat and prof. Daniel Freund, I took up the challenge of trying to estimate all locations in the city of Detroit where it would be feasible to install a curbside EV charging stations. I solved this problem by using Computer Vision to analyze the Google Street View imagery dataset, ultimately being recognized the "Most Impactful Project" award at the MIT Machine Intelligence for Manufacturing and Operations Symposium 2024.
Throughout this year working on this project, I developed an end-to-end pipeline that starts from a data extraction pipeline (with funding raised from my pitch at Google Cloud), labelling algorithms, and Vision Transformers models (Space-time attention for video understanding). To dive deeper, check out my project poster.
Video to Audio Multimodal AI
Working closely with my friends at Adorno.AI, I developed a multimodal embeddings-based end-to-end pipeline for matching sound effects to video events. This include a scene splitting algorithm, a semantic search / RAG system, and an autosync heuristic. I loved working on this project and there's much more to come. Here is me presenting this work for the course Advances in Computer Vision at the MIT CSAIL.
Urban Mobility Optimization
As part of the MIT Analytics Lab competition, my team and I partnered up with Akkodis, a leading engineering consulting company, to optimize the urban transportation system in Leuven, Belgium. We designed several multi-objective optimization models and we developed an interface to allow the municipality of Leuven to make robust decisions in such a complex and multifaceted system. We won the first place in the competition! Check out this article about us.
LLM Inference Optimization
As part of the AGI House Hackathon in April, we developed a prototype for a product that later became a startup and won the MIT $100k competition. We prototyped a proxy layer that routes LLM queries to the most efficient, maximum sized model using a caching system, a query complexity estimation and prompt optimization. In simpler terms, imagine routing the query "Hi, how are you?" to GPT2 instead of GPT4 since there is no need to use a powerful model for such a simple query. At a large scale, the cost savings and latency impact can be massive. At the hackathon, we arrived 2nd!
Other relevant work
- Instructing Reinforcement Learning Agents with Natural Language: As part of the Sensorimotor Learning class project at MIT, my team and I investigated the idea of leveraging LLMs to translate a human query to a reward function, allowing for easier interfaces between human and RL agents.
- Optimization Aware Active Learning: As a complementary project to my research work, I ideated and evaluated a novel approach to active learning (i.e. the art of deciding which data to label). This approach can be used when the ML model is an input to a optimization model and uses dual values and shadow prices to determine the most valuable data. Check out the report.
- LLM-based pre-processing framework for tabular ML: Building on top of colleagues work at the MIT ORC, my teammate and I developed a framework to use LLMs to automate some manual pre-processing tasks in the data science workflow. Check out the report.