cs @ ut austin Β Β·Β swe intern @ google Β Β·Β i turn data into weird, unexpected stories
π π π π
i find pain points across businesses, research, and people's everyday lives, then pinpoint solutions with data, machine learning, and ai.
my goal as a computer scientist is to make software and hardware IMPROVE people's physical and mental health, not just make a workflow more efficient.
i'm also someone who runs on matcha, caramel frappes, sweet treats, and meeting new people. the works.
- pulsegrid accepted 10,000/10,000 simulated device readings at 1,003 events/sec with 106.6 ms p95 request latency. it handles multi-tenant ingestion, preserves per-device order across 6 kafka partitions, retries failures, and exposes prometheus metrics.
- fleetlake processed 200,000 simulated events at a median 73,019 events/sec, then used spark to handle late data and duplicates before serving raw history through trino and fault search through elasticsearch.
pulsegrid handles the live ingestion problems: request latency, ordering, retries, and backpressure. fleetlake handles what happens after kafka: replay safety, stateful processing, storage, and querying. together, they model the path from a vehicle reading to something an engineer can investigate.
- google, swe intern
- enterprise mobility, technical pm intern: architecting an ai vehicle authorization system across 9,500+ branch locations
- afterquery (yc w25), ai data engineering intern: benchmarked generative ai tool calls across 10+ domains at 2m+ user scale
- ut austin cs dept, undergrad researcher: debugged distributed pipelines on 10m+ weather records for nasa + txson climate models
- ut austin fri, robotics researcher: built a ros2 system in c++ turning live asl gestures into physical robot motion
- reposter: cut out the repetitive part of advertising across facebook groups with a chrome extension built around the debugger api, supabase auth, postgres rls, and stripe webhooks. an older private listing supported 1,000+ paying users; the current public listing is separate.
- trialscout: turns a plain-english patient description into ranked recruiting clinical trials using cohere embed, rerank, and command.
- glassdoor alpha: explored 17k+ employee reviews as an alternative financial signal. in the apple case study, monthly senior-management ratings and stock price had a 0.376 correlation, an exploratory result rather than a trading claim.
- legacyhub: turns scattered student-organization documents, vendor history, and decisions into cited handoff answers using copilot studio, power automate, and sharepoint.
- diagnostic vlm: mapped longitudinal health records to 1,000+ abdominal ct scans for 5-year disease prediction, then tested 70 adversarial edge cases.
π π π π
- π linkedin: in/purnoor-sharma
- π email: purnoor.sharma@utexas.edu
- π github: @noorps