About
I build AI systems that actually do things, not just sit in polished demos pretending to matter. Most of my work revolves around LLMs, automation, and full-stack systems where messy ideas get turned into usable products. I’ve built tools like Deep Search and developer utilities that focus on saving time instead of creating more noise. I learn by building, breaking, and fixing systems under real constraints. That’s where most people stall. I don’t.
I care about systems that scale, don’t collapse under pressure, and don’t frustrate users. Clean architecture, practical performance, and actual usability matter more to me than hype or buzzwords. If something is repetitive, I automate it. If it’s complex, I simplify it. If it’s inefficient, I redesign it.
Sleep is for the weak. Or at least that’s what I tell myself while I’m still debugging at 3 AM.
Work Experience
- Built KRUTRIM RAG, a fully offline document-search and Q&A platform ingesting a 900+ page aerospace corpus in an air-gapped environment using Docling layout-aware extraction, conditional OCR, and heading-aware chunking preserving hierarchical section paths.
- Architected a distributed worker fleet (text, vector, graph) on a pull-based job queue with leases, heartbeats, retries, and attempt caps, running a checkpointed two-phase ingestion pipeline with per-stage resumability and crash-safe caching.
- Implemented hybrid retrieval combining Qdrant vector search, an in-house BM25 index, and a 5-layer Neo4j knowledge graph, fused via reciprocal rank fusion, Jina listwise reranking, and MMR diversity selection.
- Extended a hardware-aware multi-user load balancer routing requests by lowest estimated completion time (ECT), scored from real-time CPU/RAM/GPU (VRAM) telemetry and tokens/sec throughput, handling 10 concurrent users with up to 4 parallel requests per node.
- Optimized ingestion from 5,245s to 1,821s while maintaining 91.6% knowledge coverage via GPU-accelerated NER, five parallel embedding workers, and batched writes.
Projects
Open Source Contributions
Education
Building Under Pressure
Competed in 4+ hackathons during university, turning ideas into working prototypes in 24–48 hours alongside talented teams.
- H
Hugging Face Agents & MCP Hackathon
Online
Built intelligent AI agents and integrated the Model Context Protocol (MCP) to seamlessly connect models with external tools, APIs, and data sources. - C
CVMU Hackathon 3.0
Anand, Gujarat
Developed a social media analysis platform with real-time data integration, supporting platforms like YouTube, Instagram, and Facebook. - N
Nosu AI Hackathon
Online
A comprehensive web application integrating multiple AI features: Chat Assistant, PDF Document Analysis, and YouTube Video Summaries, built with Next.js, React, and FastAPI. - L
Level SuperMind Pre-Hackathon
Online
A web application for analyzing social media content using AI, providing insights, engagement analysis, and suggestions for improvement with a clean chat interface.