// career
Work Experience
AI Research Intern
Meta
Jan 2026 – Jun 2026
RL Fine-tuning Infrastructure for LLM Reasoning & Hallucination Reduction
- Engineered a large-scale GPU cluster training pipeline in Python with automated SLURM job submission for distributed training runs.
- Built data generation and reward-extraction components with logging, monitoring, and experiment tracking to support RL fine-tuning research targeting LLM reasoning and hallucination reduction.
- Supported curriculum-scheduled training runs and out-of-distribution (OOD) evaluation sets as part of infrastructure powering reinforcement-learning fine-tuning research for large language models at Meta.
Software Engineer II
Venwiz
Mar 2023 – Jan 2025
Search Indexing Pipeline & Hybrid Retrieval Platform
- Architected a cloud-native search indexing pipeline (AWS Lambda, SNS, SQS) decoupling SERP scraping from Elasticsearch/OpenSearch indexing, preventing scrape failures and traffic spikes from degrading search availability; used RabbitMQ for async inter-service communication.
- Designed a three-mode hybrid retrieval system for industrial procurement queries — semantic vector search, Lucene kNN/keyword scoring with weighted field relevance, and Boolean must/should/must_not matching.
- Owned Neo4j graph infrastructure for vendor-category relationships (APOC traversals, composite indexing, schema evolution); benchmarked Neo4j+Elasticsearch vs. Elasticsearch-only and consolidated architecture where recall was comparable.
- Built a full-stack self-serve taxonomy tool (Flask + Retool) with GPT-based metadata enrichment, cutting search-ops turnaround time from 3 days to under 30 minutes.
Software Engineer
Razorpay
Jul 2021 – Feb 2023
Payments Auth Path & Platform Modernization
- Reduced P99 latency 45% (820ms → 450ms) on the payments auth path serving 8M+ merchants and ₹5T+ GMV, via goroutine concurrency, database access redesign, and replica-backed reads at 15K+ TPS peak.
- Shipped Rule-Based Method Enablement, an event-driven eligibility engine processing 500K+ merchant events/day, auto-configuring payment methods via business rules.
- Led payment-terminal service decomposition, migrating 12+ read/write paths to service ownership and resolving 341 cross-module test failures while preserving transaction correctness.
- Increased merchant activation success from 91% to 99.3% by offloading reads to replica paths, cutting primary DB contention 60%; drove Grafana-based production observability and webhook alerting.
