Rahul Sharma
Research Experience
Aug 2025 – Present
Bangalore

JazzX AI logoFounding Research Engineer, JazzX AI

Reasoners for explainable AI · Advisor: Dr. Nimar Arora

  • Leading process reward modeling for due-diligence tasks, benchmarking the MAUD dataset against frontier models (Claude Opus 4.5, GPT-5) with a 17% abstention score.
  • Architected Knowledge Studio for policy management, reducing the average policy lifecycle from a week to 2 days.
  • Designed evaluation benchmarks for 200+ financial documents with 6 SMEs, using lifecycle-aware rubrics.
  • Built a v0 neurosymbolic reasoning system with a 99% accurate extraction pipeline and ~1-minute end-to-end inference.
  • Led benchmarking of GPT-5.x, Gemini 3 Pro, NuExtract, and Claude Opus for document extraction, evaluating field-level accuracy against ground-truth data and achieving over 90% accuracy.
Aug 2024 – Aug 2025
Bangalore

Microsoft Research logoResearch Fellow, Microsoft Research

LLM inference — Dr. Ashish Panwar

  • Implemented the vAttention CUDA kernel for KV-cache management in vLLM — 1.23× throughput over paged attention.
  • Added prefix caching to vAttention so queries sharing a prefix skip recomputation.

Healthcare — Dr. Mohit Jain

  • Architected an expert-in-the-loop WhatsApp bot framework handling 500+ concurrent messages/min.
  • Deployed CataractBot: 3,000+ patients, 10,000+ doctor-verified answers, 3 hospitals, 5 languages.
  • Led AshaBot with Khushi Baby, reaching 10,000+ community health workers, scaling toward 1M+.
Software Developer Experience
Feb 2024 – Aug 2024
Bangalore

Rubrik logoSoftware Developer II, Rubrik

Cloud Native Protection

  • Azure VM backup and recovery for Ultra Disk and Premium SSD v2, managing 20+ PB across 6,000+ customers.
  • Set up telemetry and dashboards in Chronosphere and Snowflake for Azure Blob backup.
Jul 2021 – Feb 2024
Bangalore

Microsoft India logoSoftware Developer II, Microsoft India

Microsoft Teams

  • Delivered a notification service handling 50M+ daily notifications with 100% reliability and EUDB compliance.
  • Cut COGS by 80%, doubled delivery speed and reduced network calls 10× via VMSS migration and API redesign.
  • Reduced Teams PowerShell module load time from 13s to 5s across 63K tenants.