Prabhjot Kaur Batth
Crew · IIT Madras

01 / Hello

Hi, I'm Prabhjot Kaur Batth.

Data Science undergrad at IIT Madras. I build AI systems, take them from idea to live deployment, and check that they actually work.

NPTEL Elite + Silver certificate, Psychology of Learning

02 / About me

About me.

Final-year student at IIT Madras. BS in Data Science. CGPA 7.36.

Prabhjot Kaur Batth

Prabhjot Kaur Batth

Final-year student · BS in Data Science, IIT Madras

CGPA 7.36 · India

Education

Jul 2023 – Present · Final year

Indian Institute of Technology Madras

BS in Data Science and Applications · CGPA 7.36 / 10

Machine learning
Machine Learning Foundations, Machine Learning Techniques, Machine Learning Practice, Tools in Data Science
AI
End-to-End AI Agent Development (A grade), LLM Security & Cybersecurity (A grade)
Programming
Programming, Data Structures & Algorithms using Python, Programming Concepts using Java, Database Management Systems, Modern Application Development I & II
Software engineering
Software Engineering, Software Testing, System Commands (Linux)
Business
Business Data Management, Business Analytics

Class 12 · Non-Medical

Akal Academy, Kajri

93%

Class 10

Akal Academy, Kajri

97%

Leadership

  • Community Library Management System: Product Manager, frontend engineer and tester in a team of 5. Top grade scorer.
  • Soft-skills project, IIT Madras: currently leading a team of 5 as team leader and designer, in Strategies for Professional Growth.

Certifications · NPTEL 2026

NPTEL Believer (Jul–Dec 2026): qualifies with 4 exams taken and all 4 certified in one semester.

QueueCare booking screen with a 3D body map

03 / The work

The work.

From smoke alerts for Indian schools to hospital queues, library systems and a kirana store's ledger.

2026 · Solo · Python, Flask, Vue 3, AWS Lambda, EventBridge, S3, CloudFormation, Gemini

UrbanWise: a child can't fix the air, so we warn them

A bad-air-day assistant for Indian cities. Most AQI apps tell you what the air is like. UrbanWise also shows where the smoke is coming from, whether more is on the way, and what to do about it.

  • Smoke Trail. Traces a city's air 48 hours back on the wind and matches the path to NASA satellite fire detections.
  • Incoming Smoke Alert. Runs today's fires forward on forecast winds and re-checks every alert on a second weather model (GFS).
  • School Mode. Go, limit or cancel for assembly, PE and dismissal. Drafts parent notices in English, Hindi and Punjabi for the principal to review.
  • Engineering. An hourly EventBridge → Lambda → S3 precompute for 8 cities. Gemini only writes the words. Every number comes from the data.

Checked against NOAA HYSPLIT.

2025 · Vue 3, Flask 3.1, JWT, Redis, Celery, SQLite

QueueCare: hospital intake, from symptom to seat

  • A "where does it hurt?" 3D body map suggests a department (knee → Orthopedics).
  • Boarding-pass tickets with QR codes and a live queue position, plus a split-flap "Now Serving" lobby board.
  • Doctors paint their weekly hours with a brush. Patients see a free-slot heatmap.

2026 · Team of 5 · FastAPI, PostgreSQL, Redis, React 19, TypeScript, Tailwind, Playwright

Community Library Management System & Reading Club

I was Product Manager, Scrum Master, frontend engineer and tester. The project got an S grade.

  • 6 roles. Catalog, borrowing, reservations, seat booking, reading club, Razorpay payments and guardian-linked child accounts, in English, Hindi and Punjabi.
  • The team built a LangGraph ReAct chatbot that calls 10 backend services as tools, with input and output guardrails.
  • I led requirements and sprints.

2026 · Team of 2, team leader · Python, Polars, LightGBM, RapidFuzz

Amazon ML Challenge 2026: Business Entity Resolution

Link business records from 3 sources to the same real business despite typos, shuffled words, names in 9 Indic scripts and broken addresses. The test set has a country (France) that never appears in training. CPU only, with no external data or pretrained models.

  • I treated it as an assignment problem: each S2/S3 record belongs to at most one S1 entity.
  • New blocking keys and a learned re-ranker raised recall@10 from 93.3% to about 98.7%.
  • Three LightGBM stages. Decoy detection gave +0.010, the biggest single jump.
  • Withheld 18% of S1 entities to simulate the test set and close the offline-vs-leaderboard gap.

Top score of 0.9861.

2024–25 · Microsoft Excel

Sales & Inventory Analysis: Mukesh Kirana Store

My Business Data Management capstone for a real grocery store in Puranpur, Uttar Pradesh. I digitised a year of handwritten invoices: 31 products, 384 product-month records, 18 data-entry errors fixed.

  • ₹59.6 L revenue, ₹4.82 L profit, 8.1% net margin.
  • April–July brings in 73.7% of yearly profit. About ₹9.7 L of stock is tied up in slow movers.
  • 17 stockouts, 6 of them Maggi.

An 8-point plan projects +43.7% annual profit (about ₹2.1 L). That's a projection, not a measured result. The owner adopted the shelf method and the Excel tracker.

Course project · Flask, SQLAlchemy, Bootstrap 5

Hospital Management System V1

Admin, doctor and patient roles. Doctors set 30-minute slots for the next 7 days. Booking takes 4 steps and the server re-checks for double booking. 39 pytest tests on GitHub Actions. QueueCare is its V2.

Smaller cargo

Kaggle practice & side repos

  • Kaggle notebooks covering regression, classification, tuning and cross-validation: 0.972, 0.7008, 0.6265.
  • streaming-llm-api: a Node/Express server that streams Groq Llama 3.3 70B over Server-Sent Events and logs first-token latency.

2026 · Python, FastAPI, LangChain/LangGraph, Prisma, PostgreSQL, React, TypeScript

Agentic Commerce & AI Safety Governance Layer

How do you give an AI agent financial agency without giving it unrestricted authority over money? I wrote the agentic commerce layer for a community-library platform (30 of 32 commits).

  • Agent-readable catalog. GET /api/v1/agent/catalog returns membership tiers, pricing, eligibility, stock signals and purchase-action templates as JSON.
  • AI upsell. An LLM reads a member's borrowing history and writes the reason for an upgrade. The server sets the price from the PricingPlan table and rejects prices sent by the client. If the LLM fails, it falls back to rules.
  • Guardian auto-pay. Pays a child account's overdue fines with zero clicks. A trust ladder raises the limit 1.2× after on-time returns and lowers it 0.7× after late ones. Hard caps: ₹200 per payment, ₹1,000 per month.
  • Bounded checkout. Proposal → approval, with a 15-minute price-lock token and mandatory human approval. Idempotent requests and atomic rollback.

04 / Process

Checked, not just claimed.

Meet the people it's for, build the smallest thing that works, check it against something real, and be honest about where it breaks.

  1. 01

    Meet the people it's for

    I build for real Indian users: a kirana store owner in Puranpur, school principals during Delhi smog, patients in hospital queues. That's why things ship in English, Hindi and Punjabi.

  2. 02

    Build the smallest thing that works

    Get one path working end to end, then deploy it. UrbanWise runs live on AWS Lambda in Mumbai.

  3. 03

    Check it against something real

    Not just a demo. UrbanWise was checked against NOAA HYSPLIT. For the Amazon ML Challenge I simulated the test distribution.

  4. 04

    Say where it holds and where it doesn't

    The UrbanWise write-up says the coarse wind grid was the weak point. The kirana profit figure is labelled as a projection.

05 / Tool kit

Tool kit.

The languages, frameworks and cloud tools I actually use, each one backed by a project.

Languages
Python, SQL, JavaScript, TypeScript, Java, C++, HTML/CSS
AI & LLM
LangGraph (ReAct agents), LangChain, tool calling, retrieval (keyword FAQ + embedding similarity), Google Gemini, OpenAI, AWS Bedrock, Ollama, Groq, guardrails, LLM fallback chains, rate limiting, streaming (SSE)
Data & ML
Pandas, Polars, NumPy, Matplotlib, Seaborn, scikit-learn, LightGBM, RapidFuzz, regression, classification, tree models, SVMs, neural networks, feature engineering, ETL, evaluation design
Backend
FastAPI, Flask, Flask-RESTful, Flask-SQLAlchemy, Flask-Login, REST API design, JWT/RBAC, Google OAuth, Razorpay, Prisma, Celery
Frontend
React 19, Vue 3, Vite, Tailwind, Bootstrap, TanStack Query, Framer Motion, i18next
Databases
PostgreSQL, SQLite, Redis, DuckDB
Cloud & DevOps
AWS (Lambda, EventBridge, S3, SSM, CloudWatch, CloudFront, CloudFormation), Docker, Docker Compose, GitHub Actions, Vercel, Git, Linux
Testing
pytest, Vitest, Playwright, Ruff, mypy, ESLint, Prettier
Business
Excel, Google Sheets, product management, Scrum, Jira, requirements gathering

06 / What I do

What I do.

AI engineeringFull-stack developmentData engineeringMachine learningBusiness analysis
  • AI engineering. LLM features with guardrails and fallbacks, where the server, not the model, owns the numbers. See the agentic commerce layer and UrbanWise.
  • Full-stack development. FastAPI or Flask backends with React or Vue frontends, tested in CI and deployed.
  • Data engineering. Scheduled pipelines like UrbanWise's hourly EventBridge → Lambda → S3 precompute.
  • Machine learning. Multi-stage LightGBM entity resolution, top score of 0.9861.
  • Business analysis. Turning a shop's handwritten invoices into decisions the owner actually used.

Looking for Software Engineering and Data Science internships: AI engineering, full-stack, data engineering and data science.

Ship toLinkedIn Scan to connect · PKB-07

07 / Say hello

Say hello.

Open to Software Engineering and Data Science internships.