I'm a CS major at Georgia Tech focusing on modeling&simulation and intelligence!
✨ Experience: Startups, NLP, Search, AWS cloud development
🎯 Goal: Work directly with customers at an AI-forward startup
⛰️ Touching Grass: I love to mountain bike, run, listen to music, and DRINK COFFEE
✨ What: Training a differentiable search index (DSI) model!
🤨 Why: Current embedding models are limited in their understanding of documents such as their inability to effectively handle negatives: "show me all papers that include x but that don't include y." There is also a relatively low ceiling on how much performance improves by increasing the number of model weights.
⚙️ How DSI's Work: A foundation LLM is "overfit" on a corpus of documents and maps them to docid's then trained on sample queries to match the docid's to questions. This way, the model learns the topics in the documents and can in turn map documents to questions and simply return docid's. Essentially documents become stored in model weights instead as vectors in the case with embeddings.
☝️ My approach: I'm testing a multistage search architecture where an embedding model is used as the first layer, being able to route a query to the best k DSI's for a given domain. Each DSI gives their ranked docid's and the corresponding docs are fed to a final reranker.
🏃➡️ Progress: In the research stage and designing the training code!
🏆 HackGT 2025: (Best use of TwelveLabs API)
Developed NewsCap: a fact checking platform designed to help users verify information and combat misinformation. It automatically analyzes YouTube videos for claims, conducts research, and presents findings to the user.
Techstack: React, CedarOS, FastAPI, Exa, TwelveLabs, Docker
🥇GaTech Hacklytics 2025: (1st in Assurant-sponsored track)
Developed Crossentropy: a version of OpenAI's Deep Research that works on local documents. Crossentropy leverages a fine-tuned version of Deepseek's R1 8B model (using GRPO) to maintain robust reasoning abilities while integrating search capabilities. The platform is built to enable users to perform natural language queries over their documents and receive iterative, contextually refined results. In addition, it supports proactive monitoring by sending real-time alerts when new, relevant documents are added.
Techstack: Deepseek R1 8B, NextJS, FAISS, Huggingface, TRL, Flask, FastAPI
🥈GaTech Hacklytics 2024: (2nd in Intel-sponsored track)
Developed Easy Deep Learning: a platform to fine-tune image classification models with minimal training data. Simply upload the photos you want to fine-tune a model on and allow us to expand the training set by over double. Utilizing a refined version of Stable Diffusion, we generate images similar to those that were uploaded and perform a number of transformations. Next, pick from a selection of classification models, helping you to select the best model for your situation. Once trained, effortlessly deploy and test your model.
Techstack: Stable Diffusion, MKL, VGG-16, Google Vision Transformer, Intel Developer Cloud, Firebase, Flask, Next.js, AWS
Spoiler: It has a lot to do with OpenSearch
1. Shipped a production-bound Route 53→ALB→Lambda path to OpenSearch Dashboards for VPC-isolated clusters, helping on-call engineers diagnose problems 2-4 hours faster per incident.
2. Developed an automated load testing pipeline for OpenSearch Clusters using Step Functions (Distributed Map), Lambda, S3, and an EC2 OpenSearch Benchmark runner.
