Erwin Deng
AI / ML @ MIT
Graduate student at MIT, focused on applying machine learning and operations research to real-world decisions, from manufacturing defect detection to wildfire resource deployment.
About
A bit about me
I'm a graduate student at the MIT Master of Business Analytics, Operations Research Center, where I work on machine learning and analytics problems that sit close to real operational decisions: assembly lines, wildfire crews, construction schedules, and financial documents.
Before MIT, I studied engineering at CentraleSupélec in France, and worked as a data scientist intern across manufacturing, luxury goods, and investment banking.
Projects
Selected work
Each card links to a full case study with the problem, approach, and impact.
Continual Prompt Optimization for Document Extraction
MIT Sloan / BMW Group
A GEPA-style closed loop that extracts, evaluates, and rewrites its own prompts. Our team raised mean extraction score on BMW repair orders from 0.33 to 0.83 without retraining any model.
Early Defect Detection on Car Assembly Lines
Coming soonMIT Sloan / Ford Motor Company
Computer vision models that catch assembly-line defects earlier, using images captured during production.
Quantifying Wildfire Crew Effects with Double ML
Coming soonMIT Operations Research Center — Research Assistant for Professor Jacquillat
A double machine learning model that quantifies how firefighting crews affect wildfire outcomes, to guide resource deployment.
Predicting Construction Project Delays
Coming soonMIT Sloan / Suffolk Construction
ML models predicting delays across 200+ construction projects, surfacing over $10M in future losses.
Universal Defect Detection for Watch Components
An anomaly detector that learns what a good watch part looks like from a few dozen images, so a new component needs no defect labeling.
Document Structure Analysis for Investment Banking
Coming soonSociete Generale — Investment Banking
Fine-tuned models for table parsing and reading order identification to improve document understanding.
Storm Damage Assessment from Satellite Imagery
EY Open Science Data Challenge 2024
A Co-DETR model that finds damaged buildings in post-hurricane satellite images. 2nd runner-up out of 11,000 entrants, presented at IEEE IGARSS 2024.
Experience
Where I've worked
Research and industry roles applying ML and analytics to real operational problems.
MIT Sloan / Ford Motor Company
Data Scientist Intern, MIT Capstone Project · Cambridge, MA
2026 – Present
- Developing AI models for early defect detection in car assembly lines using images (Python)
MIT Sloan / BMW Group
Generative AI Lab Team Member · Cambridge, MA
Spring 2026
- Built a closed-loop LLM evaluation and prompt optimization pipeline for repair-order document extraction
- Developed deterministic and LLM-as-judge evaluation framework in Python to assess structured JSON extraction quality
- Improved extraction performance by 50% over five iterations using LLM reflection and Pareto-based prompt evolution
MIT Operations Research Center
Graduate Research Assistant for Professor Jacquillat · Cambridge, MA
2025 – 2026
- Developed a double ML model to quantify crew effects on wildfires and guide resource deployment decisions (Python)
- Expanded a wildfire crew dataset by integrating satellite embeddings and weather data
MIT Sloan / Suffolk Construction
Analytics Lab Team Member · Cambridge, MA
Fall 2025
- Developed ML models to predict delays on 200+ construction projects, identifying future losses of $10M+ (Python, SQL)
Richemont
Data Scientist Intern, Research & Innovation · Buttes, Switzerland
Spring 2025
- Analyzed manually labeled datasets for watch components, identifying issues such as inconsistent labeling
- Developed universal defect detection model that works across any watch piece, enabling cost-free assessment (Python)
- Presented findings to multiple brands and delivered a production-ready solution (Python)
Societe Generale
Data Scientist Intern, Investment Banking · Paris, France
Fall 2024
- Fine-tuned AI models on domain-specific data for document structure analysis, improving accuracy for investment banking
- Developed table parsing and reading order identification models to enhance document understanding (Python)
Education
Academic background
Massachusetts Institute of Technology
2025 – August 2026
Candidate for Master of Business Analytics, Operations Research Center — GPA: 5.0/5.0 · Cambridge, MA
- Selected coursework: Machine Learning, Optimization, Advanced Analytics Edge, Hands-on Deep Learning, GenAI Lab, Power and Negotiation, Communication through Data
CentraleSupélec, Université Paris-Saclay
2022 – 2024
Bachelor and Master of Engineering — GPA: 4.0/4.0 · Paris, France
- Coursework: Advanced Statistics, Partial Differential Equations, Software Engineering, Economics, Climate Sciences
- Teaching Assistant: Led algorithmics tutorials (graphs, dynamic programming, etc.) for 30 first-year undergraduate students
- Leadership: Awarded the CentraleSupélec scholarship for involvement in Student Life and Associations
- Community involvement: Raised $15,000 and organized a 5-week mission in Nepal supporting children's education
Skills
Technical toolkit
Languages
ML & Data
Tools & Web
More
Additional information
Awards
- Bronze Medal, European Junior Olympiad in Informatics · 2017
- 1st Prize out of 220,000 in the Algoréa French Computer Science Competition · 2019
- 2nd runner-up out of 11,000 in the EY Open Science Data Challenge 2024 — see the storm damage assessment case study
Languages
- French — Native
- English — Fluent
- Chinese — Conversational
Interests
- Fencing (organized a national student competition in France)
- Piano
- Running