I'm Kundan Singh Sorout, Lead Data Scientist at HP Inc. I lead five engineers on a multi-agent system that answers Sales and Marketing questions straight out of the company's own data warehouses. Eleven years in engineering, seven in data science — vision, NLP and diffusion shipped into retail, manufacturing and ecommerce before this. On my own time I write the same stack again alone on one 6 GB laptop GPU, same open-weight models, no cluster. Building it small is how I find out which parts of a stack actually hold.
M.Tech in AI/ML from NIT Warangal. B.Tech in Computer Science from MDU Rohtak. TensorFlow- and PyTorch-certified.
Once as something a company depends on, once as something I wrote myself to see how it works. The second version is what makes me useful on the first.
Tool-augmented agents that plan, call SQL and Python in parallel, and stay coherent across a long conversation. At work that means a system five of us maintain against enterprise data warehouses. At home it means writing the orchestration loop myself to see where it frays.
Schema-aware RAG over real databases, vector retrieval, and keeping context windows lean. Most agent failures are context problems long before they are model problems — which you only learn by watching one run out of room.
Before LLMs took over there were six years of the rest of it: SDXL and GFP-GAN for a talking-head avatar, BARK for its voice, YOLOv5 detectors on manufacturing floors, VGG16 and BiLSTM across millions of SKUs — and a UNet2D + DDPM diffusion model written from scratch, because using the API teaches you nothing about the noise schedule.
Everything gets rebuilt on one 6 GB RTX 3060 before I trust it: quantization, KV-cache pressure, Ollama and vLLM, and a deep-RL trading engine over 1,400+ stocks. Neither had to exist. Both changed what I am willing to trust in a library I did not write.
An internal multi-agent chatbot connecting Sales and Marketing teams to their own data warehouses in natural language. Tool-augmented agents reason over warehouse-scale data, return charts and explain themselves. I lead the team of five that builds it.
An agentic text-to-SQL platform: RAG over database schema documentation, tool registration in a single LLM call, session memory for follow-up questions, generated Python analysis and Plotly charts. Runs on Ollama with qwen3:8b.
A deep reinforcement learning trading system covering 1,400+ stocks on a custom multi-factor environment with DQN agents. Runs entirely on local hardware — the codebase never leaves my machine, which is what pushed me into local inference in the first place.
By prompt fifty the assistant forgets your instructions, hallucinates variables and quietly triples your bill - because it re-reads the whole conversation every turn. What actually fixes it.
Read on MediumDeterministic where possible, agentic where necessary - engineering an honest job-apply pipeline that runs entirely on your laptop.
Read on MediumEvery guide assumes 12 GB, 16 GB, or a $4,000 workstation. This one is for people serious about local AI on a budget laptop with 6 GB of VRAM.
Read on MediumPrivate 1-on-1 mentorship for serious students — Data Science, LLMs, multi-agent systems and quant. Three classes a week. You read first, you explain it back, I correct. It is the method I use on myself. ₹10,000 per student per month in a batch of two.
I read every message myself and usually reply within 48 hours. Email is fastest; LinkedIn, GitHub and Medium are all live.