I'm Kundan Singh Sorout — a Lead Data Scientist and Gen AI engineer with 11 years of engineering experience and 7 years deep in Data Science. I build multi-agent systems, RAG pipelines, diffusion models and MLOps platforms that actually ship — not demos.
Lead Data Scientist · HP Inc. · Gurugram
I started as a PHP developer in 2015 and progressively moved deeper into the ML stack — from web scrapers and Flask microservices, to CNN-based image categorisation, to chatbots, to today's frontier of multi-agent LLM systems. Every layer left a fingerprint on how I build.
Today I lead a team of 5 engineers at HP Inc. building NEO, a multi-agent AI chatbot that sits between Sales & Marketing teams and HP's data warehouses. I'm a TensorFlow- and PyTorch-certified deep learning practitioner with an M.Tech in AI/ML from NIT Warangal.
HP Inc. · Gurugram
Leading a team of 5 engineers building NEO — a multi-agent AI chatbot that surfaces Sales & Marketing insights from internal data warehouses. Also shipping the External Signals Multi-Agent Sales Intelligence Platform that reads B2B client signals from D&B Hoovers and turns them into actionable recommendations for the field.
R-Systems (a Blackstone company) · Noida
Built an Avatar Video Chatbot end-to-end — Stable Diffusion XL + GFP-GAN for face generation, BARK for voice, and Llama 2 + LangChain + PEFT/LoRA for the conversational core. Shipped SkullCandy's Tunein voice assistant on the Skull-IQ device (Rasa NLU + Dialogflow + Flask + Docker on GKE). Designed a custom Diffusion image-generation pipeline (UNet2D + DDPM) and led a Dataiku CI/CD migration that lifted team productivity by 20%.
Big Oh Tech / Craterzone · Noida
Owned product categorisation models for an ecommerce catalog — hybrid CNN + RNN and VGG16 + BiLSTM architectures on millions of SKUs. Trained YOLOv5 detectors for fine-grained product part detection on manufacturing floors.
Clavax Technologies · Gurgaon
Built Flask microservices, a Rasa-powered chatbot integrated with Telegram, and an IMAP-based email-text classification system that triaged inbound mail at scale.
PhpYouth · Delhi NCR
Where it started. Django REST APIs, web scraping with BeautifulSoup & Selenium, and PHP backends — the foundational engineering muscle that still informs how I build ML systems today.
NIT Warangal · Sep 2019 — Aug 2021
MDU Rohtak · Jul 2011 — Jun 2015
Deep Learning practitioner
A short list of systems running in production today, or that ran in production at the time. Quietly, reliably, in front of real users.
An internal multi-agent chatbot that connects HP's Sales and Marketing teams to their own data warehouses through natural language. Tool-augmented agents (SQL, Python, search) reason over Snowflake-scale data, return charts, and explain themselves.
A B2B sales-intelligence platform that pulls D&B Hoovers signals on enterprise accounts, vectorises them, and serves a multi-agent reasoning layer that proposes outreach plays.
A talking-head conversational avatar — Stable Diffusion XL for visuals, GFP-GAN for face restoration, BARK for voice, Llama 2 + LangChain for dialogue, PEFT/LoRA for personality fine-tunes.
Voice assistant deployed on SkullCandy's Skull-IQ headphone platform. Rasa NLU + Dialogflow handled intents, Flask served inference, GKE handled scale.
A from-scratch UNet2D + DDPM diffusion model trained on a custom domain dataset — built to internalise how diffusion really works rather than relying on a hosted API.
An on-device CV pipeline for a connected refrigerator — inventory tracking, freshness estimation, and anomaly alerts, deployed on Azure ML.
A few of the public repos I keep on github.com/kundan121 — a mix of original projects, learning experiments, and forks I'm actively studying.
A generic agentic-AI framework — drop-in scaffold any company can use to stand up an agent stack on its own data.
Classic NLP project — text features, classical ML baselines, and a notebook walkthrough of the whole pipeline.
A working Azure ML pipeline — training, registration, and deployment with the bits you actually need to ship.
A minimal reference implementation for a binary classifier — clean code I can refer back to for new projects.
Plus active forks I'm reading from — NeMo-Agent-Toolkit (NVIDIA), open-notebook (open-source NotebookLM), MiroFish (swarm intelligence), neuralforecast, and more on the profile.
Three private classes a week. Industry-grade Data Science, LLMs, Multi-Agent Systems and Quant — taught the way I'd have wanted to learn it. This isn't about the money — it's about the kick of changing one career at a time.
I read every message myself. The fastest way to reach me is email or LinkedIn — I usually reply within 48 hours.