Akhil Nampally
B.S. Mathematics + Computer Science · Georgia Institute of Technology
Hi! I'm studying CS and Math at Georgia Tech, graduating in May 2027. My research spans world models for robot navigation at BrainML and clinical-conversation ML at the ViTAL Lab. I spent summer 2026 at Meta working on production ML data pipelines. I'm interested in building models that can learn from complex signals and work reliably beyond a demo.
Research
Current
BrainML · World Models for Robot Navigation
Building hierarchical world models for robot navigation with learned dynamics and model-based planning. I train and evaluate models on large-scale trajectory data and work on low-level control so planned routes translate into motion.
ViTAL Lab · Dr. Hyeokhyen Kwon
Building Python and PyTorch pipelines to classify autism-related patterns in multi-speaker clinical conversations. My work includes speaker-aware audio preprocessing, feature extraction, and reproducible evaluation of CNN, SVM, and ensemble models.
Past
Living Dynamical Systems Lab · Prof. Simon Sponberg
Worked on building a real-time data acquisition system for a capacitive liquid level sensor, then used Python to develop predictive models and visualizations to make sense of the sensor data. A big part of the work was automating the data logging and calibration process, which ended up cutting calibration time significantly.
Interests
- World models & robot learning — learned dynamics and planning for navigation
- Clinical and assistive ML — extracting useful signals from conversations and sensor data
- Efficient learning — self-supervised vision and lower-cost language models
Experience
Data Engineer Intern · Meta
Worked on production ML data pipelines, migrating label serving to confidence-tiered LLM outputs. Built SQL release checks and automated freshness and label-drift alerts, expanding coverage while helping the team find data incidents the same day instead of weeks later.
Projects
Hierarchical World Models for AntMaze GitHub ↗
Comparing a Director-style planner with a learned hierarchical world model on AntMaze navigation. The project tests how learned dynamics and low-level control affect whether a planned path actually reaches its goal.
Two-Tower RecSys — Two-Stage Movie Recommendations GitHub ↗
A two-stage movie recommender on MovieLens-1M: two-tower PyTorch retrieval with logistic re-ranking, evaluated on a temporal per-user split against popularity and item-KNN baselines. It improved Recall@10 by 27.5% relative over item-KNN and raised catalog coverage@10 from 9.7% to 54.6%.
LLM-SDQ — Efficient Large Language Models GitHub ↗
A project around making large language models more efficient. I implemented switchable-precision quantization on GPT-2 so different layers could run at different bit-widths, which cut GPU memory usage by 4×. I also added LoRA adapters and fine-tuned on SQuAD, seeing a solid bump in F1 score over the baseline.
PhantomShield — Real-Time Deepfake Detection GitHub ↗
A real-time deepfake detection tool built for live video calls. I trained a ResNet18-based model to detect deepfake video with 85% accuracy, then wrapped it in a cross-platform desktop app using Electron.js that shows live alerts and a risk indicator during calls.
DINO-Toads — Self-Supervised Vision for Specimen Images
Built a self-supervised vision pipeline for fine-grained specimen segmentation using patch features, PCA, and clustering. I also worked on classification in a 30,000+ image dataset and co-authored a paper on specimen detection, clustering, and orientation mapping.
EV-Arb-Tool — Odds Evaluation & Prop Finder GitHub ↗
A full-stack tool for finding positive expected value bets across sportsbooks and DFS platforms. It pulls in odds, calculates fair probabilities, and surfaces high-value opportunities automatically — processing over 10,000 props a day at its peak.
Resume
Download Resume (PDF)Contact
Best way to reach me: anampally3@gatech.edu