Projects
ML, AI & computer vision work — from industry impact at Asurion to research at Georgia Tech.
Showing 19 of 19 projects
Urban Flood Modelling
Kaggle competition entry using Graph Neural Networks for urban flood modelling and prediction.
Real-Time Assistive Sales Bot
@ Asurion
Real-time sales bot on Claude 3.7 Sonnet via AWS Bedrock Flows. 150K+ monthly live calls, personalized product offers, +100 bps sales lift.
Rental Product Recommendation System
Kaggle competition: Collaborative filtering and content-based recommendation system for rental products.
Corn Futures Climate Challenge
Kaggle competition: Climate risk features to predict corn futures. 235+ engineered features, best CFCS score: 63.79.
DocAI Platform
End-to-end ML platform on AWS with Terraform for AI-powered document processing. Multi-tenant serverless SaaS architecture.
Sales Compliance AI
@ Asurion
26 ML/LLM models for sales compliance across Verizon, AT&T & Cricket. +45% review efficiency, $1.6M annual savings. Agentic workflows, RAG pipelines, ~65K calls/month serverless architecture.
Causal Inference for Churn Reduction
@ Asurion
Causal modeling for email campaign optimization. User segmentation and cohort analysis drove 160 bps churn reduction (~$1M impact).
OCR Appliance Tag Identification
@ Asurion
Mentored 21 Georgia Tech MS students on OCR appliance tag model. 5% faster service events, ~$60K annual cost savings.
Sales Propensity Routing
@ Asurion
CatBoost propensity model routing high-value customers to top sales experts. +12% sales lift, 6% churn reduction ($15M annual value).
Artist Identification from Painting Images
@ Georgia Tech
Benchmarked 11 classifiers (SVM, CNN, Random Forest, etc.) for artist identification from paintings. Identified optimal feature-classifier combinations.
Face Detection for Photo Retention
@ Asurion
Face detection, recognition, and clustering pipeline for AT&T & Verizon Photos. +30 bps and +150 bps 30-day retention respectively.
Accelerating MRI with AI
@ IIT Madras HTIC
Deep learning for accelerated MRI reconstruction using U-Net architectures in PyTorch.
End-to-End Visual Commonsense Reasoning
@ Georgia Tech
Multimodal R2C model with visual attention & BERT for image QA. 57.6% reasoning accuracy.
Unsupervised Motion Representation Learning
@ Georgia Tech
3D CNN auto-encoder with GRU for disentangled motion embeddings. 73% top-3 precision in video action retrieval.
Video Super-Resolution
@ NIT Trichy
Undergraduate thesis: CNN auto-encoder leveraging inter-frame temporal information for video super-resolution.
Evolutionary Neural Network for Swarm Robotics
@ Universite Libre de Bruxelles
NEAT-based evolutionary controller for swarm aggregation. +31% accuracy over MLP baseline in ARGoS simulator.
Sleep Analysis using EEG
@ Analog Devices
1-D CNN for EEG sleep state classification, +18% over traditional methods. Built custom ergonomic EEG headset with Intel Edison cloud processing.
Pepper Social Assistive Robot
@ NIT Trichy
Speaker recognition (MFCC), object recognition (VGGnet), Kinect-based navigation, and speech interaction for Pepper humanoid robot.
EXOS - Robotic Hand Exoskeleton
@ NIT Trichy
Bio-inspired cable-driven hand exoskeleton. Self-fabricated design in SolidWorks with custom PID control for high-torque servos.