I'm a cloud developer at Hewlett Packard Enterprise in Bangalore. I'm most interested in machine learning systems, particularly how LLM inference is served, cached and scaled.
Research
- Hit Rate Is Not Value: a tiered KV-cache persistence system for LLM inference, and an evaluation of six eviction policies on simulated workloads and a replay of the Azure LLM inference trace. LRU wins cache hit rate but saves the least GPU compute; the paper explains why. Paper (PDF)
Selected projects
- P2P Money Transfer: UPI-style ETH payments with QR codes and an on-chain transfer log (React, Solidity)
- Scan-Mart: a grocery store you shop by photo, using a CNN trained on 131 kinds of produce (TensorFlow, Flask)
- Food Ordering System: restaurant ordering with dietary info and PayPal checkout (Spring Boot, MySQL)
- Indoor Positioning with BLE: indoor location from BLE beacon trilateration (React Native)

