AI Engineer at Maaze Underwriting, Delhi
I build AI agents that pick up the phone, read the passport and count the passengers, then keep them running in production.
Move your cursor over my name. The letters pay attention to you.
conversations handled at once by my voice and WhatsApp agents
less manual handling time after the agents went live
lower LLM inference cost from caching and batching
passenger-counting accuracy on live video at 30 FPS
Keep scrolling to move through the projects. Hover a card to run its pipeline.
Screens passports, visas and IDs and returns one explainable risk score in about two seconds.
Every scan is written to a SHA-256 hash-chained audit ledger, so no result can be quietly edited later.
Phones truck drivers, understands Hindi and Hinglish, and books technician visits on its own.
Pulls status, place, date and time slot from one sentence like “Gaadi Delhi me loading me hai, kal morning ko service kar do”.
Turns a long video into ranked vertical shorts with motion captions.
With no API key set, it falls back to its own keyword-energy scorer instead of failing.
My own MCP server that lets AI assistants manage my WhatsApp, Google account and everyday tools.
One endpoint, many tools: any MCP-compatible assistant can message, schedule and look things up for me through one typed interface.
Counts bus passengers getting on and off from a live camera stream.
Staff wearing high-visibility vests are recognised and left out of the count.
Measures footfall, queues and table turnover for a restaurant without identifying anyone.
Built to survive broken tracking: 463 shredded tracks still resolve into the 5 correct real-world events.
Plays any YouTube video as true-colour ASCII art right in your terminal.
One command, nothing else to set up: pip install ytascii-player
Maaze Underwriting Pvt Solutions, Delhi
January 2026 to now
DMAC
January 2025 to January 2026
IGNOU, New Delhi, 2023 to 2026
CGPA 8.2 out of 10
An autonomous Twilio calling agent and a multi-tenant WhatsApp agent on AWS Bedrock and LangGraph, scaled to 1,000+ concurrent conversations.
Intent parsing and function calling across rule engines, ML pipelines and agents lifted intent accuracy by 28%; caching and batching cut cost by 30%.
Pinecone paired with multi-tenant knowledge graphs: 20% better precision and 30% lower latency.
Collision prediction fusing video with g-force data at 92%, and a 10 ms gyroscope detector at 98%.
Streaming speech-to-text that joins Teams, Meet and Zoom calls for 150+ daily users.
Dockerised FastAPI services on Kubernetes with CI/CD, versioned model rollouts and A/B evaluation.
Built BizpilotAi, a Gemini-powered business assistant, and QuickGPT, a chat and image-generation app with Stripe billing, on React, Express and MongoDB.
A web scraper that exports to Google Sheets, and an MGNREGA dashboard that caches data.gov.in data in MongoDB.
Each one comes with an access badge. Drag a badge to swing it, hover it to catch the light, click it to flip it over.
Won the DS2S International Hackathon in AI and software engineering.
Click to flip backBuilt DocuSentinel for problem statement 26188 from the Ministry of Home Affairs: an AI screen for fake identity documents.
See DocuSentinelClick to flip backRecognised in a Google Developer Groups community program.
Click to flip backAI and machine learning certification from IBM.
Click to flip backAI and machine learning certification from Cisco.
Click to flip backShipped ytascii-player to PyPI. Over 100 installs in its latest month.
View on GitHubClick to flip backAlso solving on LeetCode: 89 problems, including 16 Hard, with the 50 Days Badge 2026.
Tools I use together sit close together. Hover or tap one to see its nearest neighbours.The tools I use, grouped by what they work on together.
Seven clusters: agents and LLMs, retrieval, vision, speech, backend, data and languages.
Open to AI engineering roles, freelance agent and voice AI work, and computer vision collaborations.