Work

A pioneering initiative in collaboration with ARMMAN that leverages AI and robotics to address societal challenges, particularly in healthcare.

• Shipped a multilingual LLM chatbot with Hindi audio responses for frontline workers managing high-risk pregnancies, improving access for underserved users and reducing clinician workload.
• Built a 2-tier semantic caching layer using an approved FAQ bank to serve common queries — aiming to reduce real-time LLM inference cost by ~22% while improving response latency.
• Benchmarked embedding models and tuned similarity thresholds using production replay data; implemented LLM-as-judge rubrics across 10k+ query-response pairs to improve quality and reduce failure modes.
• Designed an analytics dashboard to track 40,000+ chatbot interactions, examining user engagement patterns and HRP topic distribution to enable effective resource allocation.
• Improved production readiness via structured logging, load testing, and unit tests for critical API endpoints.

An AI-powered platform that automates the creation of presentations and summaries from various document types, enhancing productivity for professionals.

• Architected an asynchronous OCR microservice using Celery and Redis to scale document ingestion and text extraction, improving throughput by ~40%.
• Developed two full-stack applications using Next.js and FastAPI for document conversion (PowerPoint/PDF/Word) and optimised processing pipelines, reducing end-to-end processing time by ~30%.
• Enhanced chat-with-document UX by implementing a custom math expression renderer in Markdown.

A leading provider of compliance and archiving solutions for highly regulated industries, helping organisations manage digital communications across multiple channels.

• Developed Smarsh's transcriber-api using Amazon Transcribe, adding configurable settings to increase flexibility and support compliance requirements for voice communications.
• Led the development of RESTful API microservices, facilitating cross-functional team integration and enhancing scalability, which streamlined internal workflows and improved collaboration.

A Japanese tech conglomerate revolutionising online retail, fintech, and mobile services.

• Fine-tuned fashion article classification models to 88–93% accuracy to improve cross-platform product matching in a commerce workflow.
• Optimised preprocessing and training with multiprocessing, reducing training time by ~60% to iterate and ship model improvements faster.
• Improved matching precision to support better discount-deal discovery for Rakuten's cashback extension.