San Francisco Bay Area
I'm a junior at MIT studying CS, currently doing machine learning research at MIT CSAIL. I like applied ML problems where the modeling actually ships. In the past this has been through anomaly detection on Project Leo at Amazon, RAG and LLM-powered tooling on Alexa, protein language models at the Broad, and graph neural networks for climate ML at JHU APL. I enjoy working with PyTorch, training and fine-tuning, model evals, and the infra to run them at scale! Reach me at melatg [at] mit [dot] edu.
Amazon Leo
Researching LLM introspection & self-consistency @ Andreas Lab, with applications to AI transparency and alignment Prev: Integrated an ML interpretability toolkit with a remote HPC execution platform, enabling scalable activation interventions and batched model analysis on external compute resources alongside postdoc.
Alexa Communications Built an LLM-powered codegen CLI on Alexa Communications that converts Figma JSON node trees into React Native + TypeScript screens using RAG over a UI component library. Added component constraints via embedding-based retrieval and token-to-component mapping, producing screens in under 5 minutes and saving ~12 engineer-hours per screen across 10+ Alexa Comm teams. Integrated the CLI into org-wide CI/CD with Jest unit + snapshot tests and Storybook review gates, increasing test coverage from 62% to 93% and enabling rollout across three feature teams.
Uhler Lab. Trained a protein-binding motif classifier on 1.2M precomputed protein embeddings using a Vision Transformer over sequence windows, improving F1 versus internal baseline. Scaled training and ablations on an HPC cluster to enable large-batch experiments and systematic comparison of model variants for detecting phenotype-linked sequence patterns across protein groups.
Enhanced global emissions estimation models for Climate TRACE using Graph Attention Networks (GAT), increasing prediction accuracy by 17% and reducing runtime by 43% across 10k+ cities globally. Implemented an Adaptive Graph Diffusion Network (AGDN) to better model spatial dependencies between cities, enabling more robust cross-region predictions and improved generalization. Separately, helped develop a Physics-Informed Neural Network (PINN) in TensorFlow to predict sea ice movement, achieving high accuracy with implications for navigational safety through melting ice regions for the Navy. Built automated pipelines for ingesting, cleaning, and integrating heterogeneous datasets (remote sensing, traffic data, climate models).