AI engineer · M.S. at IIT Madras

Mohit Kumar

Mohit Kumar in a black jacket and an amber wraparound VR visor, with an orange light trail streaking off behind his head

I build generative AI systems end to end: retrieval pipelines, LLM agents and fine-tuned models, plus the infrastructure that runs them.

Research and projects

Selected work

? assembly graph · edges = mate constraints retrieve library · 234,299 parts #1 #2 #3 #4 #5 #6 ⋮ median

Research · Computers & Graphics, 3DOR 2026 special section

SLOT-GNNA hierarchical graph neural network for finding the missing part in a CAD assembly

Median rank 5 of 234,299 parts · 35% better than the strongest baseline

Given an assembly with a part missing, SLOT-GNN retrieves the right part from a library of 234,299. It encodes each part's boundary representation as a graph and passes edge-conditioned messages over the assembly's mate constraints.

  • Graph neural networks
  • B-rep encoder
  • Message passing
  • Retrieval
agenttransfer 9 checks caps · anomaly pays fromyour bank held forapproval kill switch · audit log

Agentic payments · Early access

ndStill

Nine checks on every agent transfer, before any money moves

A non-custodial control plane for AI agent spending. Agents pay from a business's own bank accounts, and each one gets caps instead of a balance: per-transfer, daily and monthly limits. Transfers that are over a threshold or look unusual wait for a person to approve them, a kill switch freezes one agent or the whole organisation, and every step goes to an append-only audit log. Agents connect through a REST API or an MCP server.

  • Rust
  • PostgreSQL
  • MCP server
  • REST API
  • Next.js
› prompt live Vite + React 19 app EC2 worker per project

Generative AI · Developer tools

AI App Builder

Natural-language prompt → live React 19 app

Turns a prompt into a running Vite and React 19 app in TypeScript and Tailwind, with an in-browser code editor, project preview URLs and iterative edits. It runs on three tiers (Next.js 15, Express and Prisma) with AWS EC2 Auto Scaling workers. Each project gets its own worker, which is released when it goes idle to keep compute costs down.

  • Next.js 15
  • Express
  • Prisma
  • AWS EC2 Auto Scaling
  • Docker Compose
  • Caddy
dataDVC trainXGBoost registryMLflow serveFastAPI drift detected → retrain · Airflow + SHAP

MLOps · End to end

Credit card fraud detection

Drift detection and feedback-driven retraining, automated

An XGBoost model served through FastAPI, with DVC for data versioning, MLflow for experiment tracking and registry promotion, SHAP explanations and feedback capture. Airflow handles drift detection and retraining. GitHub Actions, Prometheus, Grafana and AlertManager cover CI and monitoring, and it deploys on a load-balanced Docker Swarm.

  • XGBoost
  • FastAPI
  • DVC
  • MLflow
  • Airflow
  • Prometheus
  • Docker Swarm
regional workers every 30 s RedisStreams ClickHouseviews p50 / p95

Distributed systems · Monitoring

Uptime Watch

Website checks every 30 s from regional workers

Regional workers check websites every 30 seconds and deliver results over Redis Streams with acknowledge-after-write. Results are stored in ClickHouse, where materialized views compute hourly uptime, median and 95th-percentile latency, and incidents. These are served through OpenAPI REST APIs and a Next.js dashboard.

  • Redis Streams
  • ClickHouse
  • PostgreSQL
  • Next.js
  • Kubernetes
  • GitHub Actions

Internships, most recent first

Experience

  1. Jun – Aug 2026

    AI Intern

    ORIGINN · startup discovery, hiring and pre-order platform

    • Built a generative AI screening pipeline that evaluated 120+ startup profiles and cut manual review from 40 minutes to 5 per startup, about 85% less time.
    • Designed a startup validation and scoring framework covering 55+ parameters across six sectors.
    • Reached about 82% agreement with founder-reviewed assessments across 30 cases by iterating on 15+ prompts against a structured evaluation set, which cut inconsistent outputs by about 40%.
    • Built an AI hiring agent that screens and ranks applications for first-pass shortlisting, and 12 REST API endpoints in TypeScript and Node.js behind the React frontend.
  2. Mar – Jun 2026

    Agentic AI Developer Intern

    Indobase.in · Terrarare Solutions Pvt. Ltd.

    • Fine-tuned a pretrained Phi-family language model with LoRA, PEFT and TRL SFTTrainer, deployed it through Ollama as GGUF, and A/B tested it against the base model.
    • Built a multi-agent editorial pipeline with planner and writer agents, a human approval gate and WordPress publishing, supporting two posts a week without manual formatting.
    • Built RAG assistants over product and policy documentation.
    • Deployed and ran a SaaS platform on a VPS with Docker, Traefik and Kong API gateway, fixing routing, CORS, JWT authentication, container networking and CI issues.
  3. Feb – May 2026

    Process Engineering Intern

    Uno Minda Ltd.

    • Helped design and set up a new automotive headlamp assembly line (YED model), from production goals and layout planning to equipment selection and test runs.
    • Helped design 12+ workstations and the material handling flow, and wrote standard work procedures and quality checkpoints for operator training and line handover.

Leadership

  • Founding team member, Product, at ndStill: banking and spend-control infrastructure for AI agents. IIT Madras pre-incubation.
  • AI Lead at Kognigen: IIT Madras pre-incubation.
  • Placement Coordinator, Department of Engineering Design, IIT Madras, Jul 2026 to present. Deputy Coordinator, Aug 2025 to Jul 2026.
  • Teaching Assistant, Data Science: Theory and Practice, IIT Madras, Jan – May 2026.

Education, skills and certifications

About

I'm an AI engineer who builds generative AI applications end to end, from retrieval pipelines and LLM agents to the APIs, cloud infrastructure and monitoring behind them.

I'm doing an M.S. in Engineering Design at IIT Madras, where my research on graph neural networks for understanding CAD assemblies was published in Computers & Graphics. Before that I did a B.Tech in mechanical engineering and robotics.

Recently I've built GenAI screening and hiring agents at ORIGINN and fine-tuned a Phi-family model with LoRA at Indobase. I also completed the 3D Vision Summer School (3DVSS) 2026, and I hold the AWS Certified AI Practitioner and AWS Certified Cloud Practitioner certifications.

  • M.S., Engineering DesignIndian Institute of Technology MadrasCGPA 8.66/10
    2024 to present
  • B.Tech., Mechanical Engineering + RoboticsDCR University of Science & Technology, MurthalCGPA 8.05/10
    2020–24

CourseworkFoundation of Machine Learning · Introduction to Deep Learning · Data Science: Theory and Practice · MLOps

AI
Large language models, prompt engineering and evaluation, RAG, hybrid and dense retrieval, embeddings, LangGraph, multi-agent systems, LoRA, PEFT, TRL SFTTrainer, Ollama, GGUF
Machine learning
PyTorch, scikit-learn, XGBoost, Hugging Face Transformers, graph neural networks, CNNs, ResNet, vision transformers, sequence models, attention, feature engineering, model evaluation
Deployment and MLOps
Docker, Kubernetes, AWS, GitHub Actions, Git, Linux, MLflow, DVC, Apache Airflow
Monitoring and data
Prometheus, Grafana, AlertManager, Redis Streams, ClickHouse, SHAP, Weights & Biases
Backend and full stack
APIs, Node.js, Express, React, Next.js, Postgres, MongoDB, Prisma, Redis
Programming
Python, SQL, TypeScript, JavaScript, Bash