Fine-Tuning Llama 3.2–3B on IPCC Climate Reports
Finetuning, Pretraining, and building a RAG pipeline using a Llama3.2-3B model on IPCC Climate Reports in a distributed environment
Portfolio
AI as a Service – Containerized ML Pipeline on Kubernetes
Cloud-native ML pipeline with CNN training as a K8s batch job and scalable multi-replica FastAPI inference on GKE
Market Recovery Dynamics During the 2020 Crash
Large-scale data lake pipeline integrating stock, COVID, mobility, and macroeconomic data via HDFS, MapReduce, and Hive/Trino for crash/recovery analysis
MOT-TR: DETR Fine-Tuning for Moved-Object Detection
Robust, reproducible pipeline designed to fine-tune the DETR (DEtection TRansformer) model for the task of detecting moved objects in pairs of images taking from parking lots and intersections (VIRAT dataset).
Limit Order Book Midpoint prediction for Cryptocurrency Stocks
Evaluated classical ML model performance and designed light-weight neural networks with custom loss functions for SoTA limit order book midpoint prediction for a cryptocurrency stock
OS Simulation Suite
Compact suite of operating system simulations in modern C++ spanning - linking/loading, CPU scheduling, memory paging, and disk I/O scheduling
CoBaLI – Continuous Batching for LLM Inference
C++/CUDA inference engine achieving 6.2× throughput via GPU-assisted scheduling on top of llama.cpp, without modifying model kernels
EcoScan – Sustainable Fashion Scanner
1st Place HackNYU 2025 (MLH) — Mobile app that scans clothing items and returns Eco-Scores with AI-powered greener alternatives
RepoRecSys – GitHub Repository Recommendation System
Two-tower neural recommender with cold-start handling, hot model reloading, and containerized FastAPI inference (<100ms latency)
VisaPulse – AI-Powered F-1 Visa Mock Interview Platform
LangGraph-orchestrated agentic interview system with voice interaction, document analysis, RAG over Reddit cases, and personalized PDF evaluations
Handwritten Character Classification Pipeline
ResNet50V2-based EMNIST classifier achieving 93.55% test accuracy with mixed precision and optimized data pipeline