USC MS Computer Science student with over two years building backend and AI software at IDFC FIRST Bank. I built agent tools with the engineering team there, and I enjoy turning unfamiliar problems into working software. Most recently, I built Drigil with a three-person team, earning 2nd place in Anthropic's Claude for Public Benefit track at Cal Hacks 13.0.
Projects
Drigil 2nd · Anthropic's Claude for Public Benefit track · Cal Hacks 13.0
Smart-glasses job coaching in the supervisor's own voice · Oct 2026
Built smart-glasses job coaching for people with intellectual and developmental disabilities in a three-person team over 36 hours.
Combined demonstration keyframes and word-level transcripts into a structured task spec with Claude; confidence-gated vision checks and cached coaching audio.
Turns one narrated demonstration into a step-by-step task specification: keyframes where hands pause, aligned with a word-level transcript.
Live coaching: confidence-gated vision checks against the camera, coaching audio cached per line in a cloned supervisor voice, with prompts that fade as the worker gains mastery.
When a step keeps failing, a sandboxed remediation agent rewrites it; a supervisor reviews before-and-after and approves before it goes live.
AI policy review with cited evidence and human approval · 2026 – present
Checks internal policies against regulatory requirements, retrieves evidence from 25,000+ documents and flags gaps with citations.
Four LangGraph agents with a 500+ query eval harness: 92% recall@5, 96% citation accuracy.
PythonLangGraphLangChainpgvectorPostgreSQL
Remediation is routed through human approval before anything changes.
Working toward Palantir Foundry / AIP ontology integration (in progress).
live · retrieve → compare → cite
NFTGenesis IEEE InC4 2024
Generative images with tamper-evident hidden signatures · Sep 2024
Fine-tuned Stable Diffusion with QLoRA in a four-person team, cutting trainable parameters by over 90%.
Paired encoder/decoder CNNs embed tamper-evident signatures with 98% verification accuracy.
PyTorchStable DiffusionQLoRACNNs
interactive · simulation
Authored the IEEE InC4 2024 paper with the team: “NFTGenesis: An NFT Generation and Authentication System with Market Intelligence Using Deep Learning Based Steganography.”
Mend Finalist · Top 15 · USC Origin Hackathon 2026
One photo becomes a verified street-repair report · 5-person team · Sep 2026
Built Mend's surge pricing: each H3 hexagon (about a city block) inside a city bounty gets a live 1×–5× multiplier that rises where coverage is thin or stale and cools as reports pile up, so the crowd spreads out instead of crowding one corner.
Wrote the reward rate card (severity × asset tier × first finder × zone, with a daily cap), itemized in-app receipts, bounty budgets, and pay-per-visit sponsored campaigns with geofenced check-ins and one-time till codes.
One shared pricing module prices both the heat map and the payout, and each report is priced before it's saved, so the number a player sees on the map is the number they're paid, and a report can't cool its own cell.
Bounties can carry a points budget and stop heating the map once it's spent, so buyer spend stays predictable. 100 points = $1.
Sponsored campaigns: a brand pays per verified store visit. A player files a real report within 300 m of a store, checks in within 75 m, and gets a code that the store redeems once from the dashboard.
Rebuilt the work on top of teammates' gift-card redemption, danger-zone and first-finder rules, which pay nothing inside danger zones and 40% for confirmations, keeping those rules unchanged.
Covered by 11 Deno unit tests for pricing and campaigns, plus 39 end-to-end checks that run every database migration and the demo seed in PGlite with PostGIS.
Team context: the five of us shipped an iOS app (SwiftUI) with on-device damage models, server-side vision verification and a city work-queue dashboard, backed by a 51-person survey and an analysis of 1.77M LA 311 requests.
Jul 2024 – Aug 2026 · previously intern, Jan – Jun 2024
Built a Go MCP server with the engineering team, choosing a shared tool layer over per-agent integrations for typed schemas and centralized auth; cut CRM and asset lookups from minutes to under 5 seconds.
Demoed loan and asset lookup tools to product and business teams, explained limits and failure cases, and was recognized for the demo at a Quarterly Review.
Prototyped LLM-based cheque field extraction with MiniMax M2.5 for compliance-document automation.
Developed Go and Python microservices, routed REST APIs through Kong with auth and rate limiting, and added Kafka retry handling.
Tuned SQL queries and indexes over 50M+ records, cutting API response times by 42%.
Set up service observability with monitoring and failure tracing, helping cut incident resolution time by 35%.
AI Research Intern
ElementalsAI · Bangalore, India
Jun 2023 – Aug 2023
Built and evaluated multimodal LLM applications with Hugging Face, scikit-learn and the OpenAI API; benchmarked foundation models.
Explained AI concepts through a 12-module curriculum with 25+ exercises.
5M+customers on the banking systems I worked on
<5sCRM and asset lookups, down from minutes
42%faster API responses over 50M+ records
35%faster incident resolution
Education
MS, Computer Science
University of Southern California · Los Angeles
Expected May 2028 · Analysis of Algorithms, Information Retrieval and Web Search Engines