Madhur Kumar Rana

Software engineer. AI/ML and data science.

I build full-stack web apps and machine-learning systems, and I ship them with tests, CI and live demos you can open and try.

Projects

Each one is public on GitHub with a README that lists its limits, and most have a live demo.

Signal Clone chat screen with a conversation list and an open chat
Signal Clone on a phone-sized screen
Full-stack web app

Signal Clone

A Signal-style messenger with real-time one-to-one and group chat.

  • Read receipts, typing indicators, reactions, replies, attachments and disappearing messages over WebSockets.
  • JWT sessions shared by the REST API and the WebSocket connection; an 8-table SQLModel schema.
  • pytest, Vitest and Playwright tests run in GitHub Actions. Frontend on Vercel, API on Render.
  • Next.js
  • TypeScript
  • FastAPI
  • SQLModel
  • WebSockets

The API runs on a free plan, so the first load after a quiet period can take up to 30 seconds.

BrowserPilot interface with example instructions and safety settings

BrowserPilot

Control a real browser with plain-English instructions. An LLM (OpenAI or a local Ollama model) plans the steps and drives Playwright through the Model Context Protocol. Guardrails block private and cloud-metadata addresses before the agent starts.

  • Python
  • MCP
  • Playwright
  • Streamlit
Terminal showing the SprintPilot offline demo report

SprintPilot

Sprint planning and Agile reports for Jira teams, run from PowerShell. Nine workflows, from backlog triage to retrospectives. Jira only changes when you pass an explicit flag, and an offline demo runs without Jira or an account.

  • PowerShell
  • Jira
  • MCP
  • CI
YOLO Object Detection Studio detecting a bus and people with a region of interest

YOLO Object Detection Studio

Detect and count objects in images, video and a live camera. Limit counting to a region of interest, track objects for unique counts and export CSV. Choose between COCO (80 classes) and Open Images (600 classes) models, filter by class and tune the confidence.

  • Python
  • Ultralytics YOLO
  • OpenCV
  • Streamlit
  • Docker
GraphCite answering a two-hop question with checked citations

GraphCite

Multi-hop question answering over a knowledge graph. Every relation keeps its source sentence, and each sentence of an answer is checked for a citation and for word-level support in it. Unsupported statements are flagged.

  • Python
  • Neo4j
  • Streamlit
  • LLM APIs
Glaucoma model output: optic disc and cup outlines and a heat map of where the network looked
Machine-learning research, ongoing

Glaucoma detection from retina photos

My M.Tech research at MNNIT. Deep-learning models outline the optic disc, cup and atrophy, measure the cup-to-disc ratio and return a glaucoma probability. The models run in the browser, so photos are never uploaded.

MeasureClassicalMine
Disc / cup Dice0.771 / 0.4480.978 / 0.904
Cup-to-disc ratio error0.1780.031
Glaucoma AUC (411 eyes)0.6020.877
  • PyTorch
  • U-Net
  • ONNX Runtime Web
  • scikit-learn

A research prototype, not a medical device.

Skills

Tools I have used in the projects above.

Software engineering

  • Python
  • TypeScript
  • JavaScript
  • SQL
  • HTML and CSS
  • FastAPI
  • Next.js
  • SQLModel
  • REST APIs
  • WebSockets
  • JWT authentication
  • Git
  • GitHub Actions
  • Docker
  • pytest
  • Playwright

AI and machine learning

  • PyTorch
  • OpenCV
  • Ultralytics YOLO
  • Deep learning
  • Computer vision
  • Retrieval-augmented generation
  • Model Context Protocol
  • Neo4j and knowledge graphs
  • OpenAI and Ollama APIs

Data science

  • scikit-learn
  • Cross-validation
  • Calibration
  • ROC and AUC
  • DeLong test
  • Dice and segmentation metrics
  • SQL

Education

M.Tech., Artificial Intelligence and Data Science
Motilal Nehru National Institute of Technology (MNNIT) Allahabad, 2025 to present

B.Tech., Information Technology
KIET Group of Institutions (AKTU), Ghaziabad, 2024

How I work

I write tests, set up CI, and document what works and what does not. The README of each project lists its limits as well as its features.

Looking for a software engineer? Let's talk.

I am open to software engineering, AI/ML and data science roles.