Kolkata · 2026

Hi, I'm Pratz

CS undergrad. Full-stack builder. I make things that work and things that feel good.

View My Work Get In Touch
Download Resume
terminal
$ whoami
> pratz — full-stack dev
$ ls ./skills
> react node django java python
$
Scroll
01 — About

A bit about myself

4th sem IT student from Kolkata. I build full-stack applications — backends that scale, frontends that look good.

Recently: built a KIVI-inspired INT8 KV cache quantization engine in PyTorch (reducing GPT-2 memory footprint by 68.5%), engineered a 5-agent AI pipeline for clinical trial reasoning (TrialMatch AI), built a voice-first "Ambient Intelligence" system for retail stores (DataTalk AI) which ranked in top 50 nationwide, and co-built a CTF site for IEEE TechWeek.

PyTorch· React· Node.js· Django· Python· Java· Git· TailwindCSS· Express·
02 — Skills

What I work with

Languages

  • Python
  • Java
  • JavaScript
  • C

AI / ML

  • PyTorch
  • Hugging Face
  • LLM Optimization

Backend

  • Node.js / Express
  • Django / DRF
  • REST APIs

Frontend

  • React
  • Next.js
  • Tailwind CSS
  • HTML / CSS

Tools

  • Git / GitHub
  • Azure
  • Vercel / Render
03 — Coursework

What I've studied

Key courses from my CS undergrad that shaped the way I build

CS Data Structures & Algorithms Sem 3
CS Object-Oriented Programming Sem 3
CS Database Management Systems Sem 4
CS Operating Systems Sem 4
CS Computer Networks Sem 3
MATH Discrete Mathematics Sem 3
04 — Projects

Featured work

A selection of projects that showcase my skills and creativity

01
Python PyTorch Hugging Face LLM Optimization

KV Cache Quantization

Python · PyTorch · Hugging Face Transformers · GPT-2 Optimization

Architected a custom inference engine in PyTorch with a KIVI-inspired INT8 quantization pipeline, bypassing model.generate() for manual token-by-token cache management. Reduced GPT-2 KV cache footprint by 68.5% with only a 0.71% increase in perplexity, implementing flexible symmetric/asymmetric quantization schemes across whole-tensor, per-channel, and per-token granularities.

GitHub Read Blog
02
AI Multi-Agent Full-Stack Healthcare

TrialMatch AI

Next.js · Express · OpenAI SDK · Groq/Llama 4 · ClinicalTrials.gov API

Multi-agent AI pipeline that takes a patient summary and a clinical trial ID, then produces a structured, criterion-by-criterion eligibility assessment. Built a 5-agent sequential pipeline (Extractor → Fetcher → Parser → Reasoner → Briefer), with Groq-hosted Llama 4 Scout for LLM inference and guardrails for JSON validation, missing-criteria backfill, and fallback brief generation.

Source Code Live Site
03
AI Multimodal Ambient Intelligence

DataTalk AI

React · Express · Azure OpenAI · Computer Vision · Speech SDK

Voice-first "Ambient Intelligence" system for Kirana stores. It captures shop activity via voice + camera feed, converting real-world retail interactions into structured business intelligence. Automates billing, generates real-time insights, and tracks inventory with zero manual data entry. Built with Azure Multimodal Fusion as part of Team RunTime Terror.

Source Code Live Demo
04
CTF Interactive Web Team

IEEE TechWeek — The Last Broadcast

Vanilla JS · CTF · IEEE TechWeek 2025 · Team of 5

Co-built a multi-challenge CTF event site for IEEE Manipal's TechWeek as part of a 5-person team. I implemented multiple clues including the dystopian bookshelf puzzle (an immersive drag-drop challenge) and the audio steganography clue, where hidden messages were encoded in spectrograms and surfaced only through developer tools.

Live Site
05 — Achievements

Recognition

Microsoft

AI Unlocked – Campus Edition

Selected for the Top 50 Prototype League out of 300+ teams nationwide with DataTalk AI.

2025
06 — Contact

Let's connect

I'm open to internships, collaborations, and interesting problems.

Mail me!