AI Engineer at Maaze, Delhi
I build RAG systems and AI agents that answer the phone, route every request to the right model, and keep latency and cost down at scale.
Move your cursor over my name. The letters pay attention to you.
concurrent conversations across my voice and WhatsApp agents
less manual handling time once the agents went live
lower inference cost from token optimisation 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.
Answers customer questions from a retrieval-augmented knowledge base, right inside WhatsApp.
Delivered natively through the Meta WhatsApp Cloud API, so customers never leave the chat they already use.
Tracks people across video frames and keeps the same ID for each person.
Appearance feature matching re-attaches the right ID when someone leaves the frame and comes back.
Ask about a fleet in plain language and get live tracking, idle time and route answers in chat.
Surfaces live tracking, idle detection, route analysis and driver-behaviour insights without opening a dashboard.
Detects vehicle collisions in dashcam footage.
Runs on any video from the command line, with the model downloaded once and kept locally.
Finds businesses, enriches their data and exports ready-to-use lead lists.
Enrichment pipelines feed exportable intelligence dashboards instead of hand-built spreadsheets.
Renders personalised try-on images by overlaying garments on a person’s photo.
Automated image pipelines turn a product photo and a user photo into a try-on visualisation.
A 3D top-down survival shooter, built and shipped in Unity.
Owned it from architecture to final build: vector movement, custom physics and pathfinding enemies.
Maaze
March 2026 to now
Crenovent
September 2025 to March 2026
Decision Point
April 2024 to January 2025
Independent
April 2022 to March 2024
Maharaja Ganga Singh University, Sardarshahar
2019 to 2022
An autonomous Twilio voice agent and a multi-tenant WhatsApp chatbot on AWS Bedrock, scaled to 1,000+ concurrent conversations and cutting manual handling time 45%.
Intent parsing sends each request to rule engines, hybrid ML pipelines or LLM agents: 28% better intent accuracy, and 30% lower inference cost from token optimisation.
Pinecone vector search paired with multi-tenant knowledge graphs: 20% better precision and 30% lower latency.
Real-time audio processing that joins Teams, Meet and Zoom calls for 150+ daily users at 92% satisfaction.
YOLOv8x with BoxMOT counts passengers at 30 FPS with 96% accuracy; collision prediction at 92%; and MOD Pizza store heatmaps that lifted campaign engagement 18%.
An enterprise agent catalog of policy-safe, one-click automations, with human-in-the-loop overrides, audit logs, evidence packs, and strict tenant isolation with RBAC.
End-to-end query latency down 20–30%, and a GPTCache layer that skipped ~35% of redundant model calls. Serverless ML on Azure Functions for event-driven retraining.
Two years building Deadtoy Showdown in Unity and C#, from architecture to final build.
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.
Stanford Online, 2025: neural networks, decision trees and practical advice for applying machine learning.
Click to flip backIBM, 2025: attention, transformer architectures and language model training.
Click to flip backIBM, 2025: applying generative AI across the data science workflow.
Click to flip backGoogle, 2025: compute, networking and storage on Google Cloud.
Click to flip backTools 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: LLMs and agents, retrieval, ML and vision, voice and chat, backend, data and cloud, and languages.
Open to AI engineering roles, RAG and agent platform work, and computer vision collaborations.