Portrait of Pritipratyush Mohapatra

AI/ML Developer · B.Tech CSE (AI & ML), OUTR '27

Hi, I'm Pritipratyush (Pupun) Mohapatra.

I build things from scratch: custom PyTorch autograd engines, tokenizers, a 10M-parameter Transformer trained in my bedroom, and apps people actually use. Importing libraries feels like cheating. (Mostly kidding.)

  • AIR 5690 GATE 2026 (DS & AI)
  • 9.05 CGPA
  • AI Intern Yavda Analytics

Things I've built

LLMs from scratch

Echoes of Meaning

A character-level GPT that learns language from chaos.

A 10M-parameter decoder-only Transformer (6 layers, 8 heads, 256-dim) with its own multi-head attention and character tokenizer, no abstraction libraries. Trained on a 4.8M-character literary corpus on a 6 GB RTX 4050 using gradient accumulation and cosine annealing.

  • 1.50 val loss
  • 4.48 val perplexity
  • 57.7% next-char accuracy
  • PyTorch
  • CUDA
  • AdamW
Privacy-first AI

secureLens

See the invisible.

A Chrome extension that turns your password field into a live heatmap of predictability. A trigram model trained on 14.3M leaked passwords, with Laplace smoothing and Shannon entropy mapped to crack-time estimates. Zero dependencies, and nothing leaves your browser.

  • 267K probability table
  • 3.8× smaller (0.74 MB)
  • <2 ms inference
  • Manifest V3
  • JavaScript
  • N-grams
NLP

Context-Aware Spell Checker

Catches real-word errors like "going to the see" with a bigram language model, Laplace smoothing and Levenshtein distance. Every fix comes with a confidence score.

  • Python
  • Streamlit
NLP pipeline

Patrakaar.AI

Summarizes news articles, extracts keywords and classifies topics with DistilBART and KeyBERT, behind FastAPI and a Streamlit dashboard. Lazy model loading made inference 6× faster and ended CUDA out-of-memory crashes.

  • Hugging Face
  • FastAPI
  • Streamlit
From first principles

Grad Engine

A scalar-valued reverse-mode autodiff engine, with neurons, layers and MLPs built on top of it. Backpropagation without a framework in sight.

  • Python
  • Calculus

Experience & education

Experience

  1. May – Jul 2026 · Hybrid, Bhubaneswar

    AI Intern, Agentic AI & ML Systems

    Yavda Analytics Pvt. Ltd.

    • Built LLM agent workflows with tool calling and RAG that automate multi-step document and data processing, replacing manual review.
    • Turned subjective agent-output checks into measurable pass/fail regression tests against baseline models.
    • Built stages of a FastAPI data pipeline that turns messy CSV/XLSX files into schema-validated records with row-level errors, and moved heavy processing to Redis-backed background workers so the API stays fast under large uploads.
    • Shipped with Docker and GitHub Actions to Google Cloud Run: lint and test gates, one-push deploys, JWT/OAuth auth.
  2. 2026 – Present

    Technical Mentor, Algorithmic Foundations

    Peer Mentorship Program, OUTR

    • Designed the curriculum and lead weekly sessions for 10+ peers on dynamic programming, sliding windows, two pointers and complexity analysis.

Achievements

  1. 2026

    All India Rank 5690, GATE (Data Science & AI)

    National exam on ML theory and applied mathematics

  2. School

    Gold Medalist, Mathematics Olympiad

    State and district level

Education

  1. Aug 2023 – May 2027

    B.Tech, Computer Science (AI & ML)

    Odisha University of Technology and Research (OUTR) · 9.05 CGPA

  2. 2020 – 2022

    Class 12, CBSE (Science)

    ODM Public School · 90%

  3. 2012 – 2020

    Class 10, CBSE

    Kendriya Vidyalaya · 96.25%

Certifications

Every one links to its verification page.

Anthropic · Jul 2026

Stanford & DeepLearning.AI · Coursera

Writing

I explain the ideas behind ML systems in plain words on Substack.

Chart placing movies on two axes by tone and theme, showing a search landing closest to 3 Idiots

How embedding models turn text into vectors, why cosine similarity finds results by meaning rather than exact words, and when you still need a traditional database.

"Meaning helps you find the candidates. Exact data and rules help you choose the right answer."

All posts on Substack ↗

The tech arsenal

Haters gonna hate, but my validation loss is steadily decreasing.

  • Python
  • C++
  • PyTorch internals
  • Transformer architecture
  • CUDA-aware training
  • Tokenizers
  • Autograd
  • NLP
  • LLMs
  • Agentic AI
  • QLoRA
  • Hugging Face
  • RAG
  • Tool-calling agents
  • XGBoost
  • Random Forest
  • Pandas
  • Laplace smoothing
  • SQL
  • Bash
  • TypeScript
  • JavaScript
  • React
  • FastAPI
  • Docker
  • Google Cloud Run
  • GitHub Actions
  • Redis
  • PostgreSQL
  • Streamlit
  • Firebase
  • Cloudflare Workers
  • MCP servers
  • Manifest V3
  • Tailwind

Let's connect

Data tells better stories than words, but a beautifully tuned loss histogram is basically a romance novel.