Career
Work Experience
AI Research Intern
Jan 2026 – Present
FactCurriculum: Adaptive Verifiable Factual Training to Reduce Hallucinations in LLMs
- Designing FactCurriculum, an adaptive auto-verifiable factual training framework for LLMs using KB-backed QA tasks and Reinforcement Learning to study hallucination reduction and confidence calibration.
- Building an evaluation pipeline to measure factual accuracy, hallucination rate, and Expected Calibration Error (ECE) against SFT baselines on 1B-scale models.
Software Engineer II
Venwiz
Mar 2023 – Jan 2025
Category Connect & Search Relevance Platform
- Designed an event-driven ingestion pipeline (Lambda + SQS + SNS) to process 0.7–1M vendor records, choosing async fan-out to handle bursty workloads and avoid backpressure on downstream search indexing.
- Built a hybrid search system in Elasticsearch combining OpenAI embedding-based semantic search with Lucene kNN and keyword relevance scoring, improving coverage and relevance for category-based vendor discovery.
- Solely owned the Neo4j graph database with complex APOC-based traversals, efficient indexing, and schema customization for industrial procurement search.
- Designed relevance signals and ranking strategies achieving 70% relevance across production queries; built a self-serve graph and tag management system using Python Flask + Retool.
- Integrated GPT-based APIs to auto-generate and enrich vendor metadata, strengthening discovery and downstream search relevance.
Software Engineer
Razorpay
Jul 2021 – Feb 2023
Payment Systems & Platform Modernization
- Improved high-throughput payment and merchant activation flows via goroutine-based concurrency and replica database configurations, reducing P99 latency by 45% and support tickets by 35%.
- Designed and productionized Rule-Based Method Enablement, a real-time event-driven system that automatically configures merchant payment methods based on dynamic business parameters using Laravel events.
- Led monolith-to-microservices API decomposition, resolving data mismatches and fixing 341 failing unit tests to stabilize releases and improve service isolation.
