Debugging life one bug at a time @ BIT Mesra

I'm passionate about applying machine learning to solve real-world problems, with a strong focus on Deep Learning, Computer Vision, and NLP. I specialize in building efficient AI systems through model optimization, parameter-efficient fine-tuning, and production-ready ML pipelines.

Occasionally debugging neural networks, datasets, and my own assumptions.

I promise I respond faster than my model's training time 🚀

🏢 Founding Member — ML & Product @ HireBuddy (Live) | May 2025 – Nov 2025

• Built and deployed an AI-powered resume-job matching system from 0→1 using RAG pipelines and Transformers, improving shortlisting accuracy by 35% and processing 10K+ resumes/day
• Designed end-to-end NLP pipelines for job-fit scoring, candidate ranking, and automated feedback generation using LangChain, achieving 92% model performance
• Deployed ML models via Flask + GCP with Docker containerization; integrated REST APIs with recruiter-facing product, increasing user engagement by 40%
• Collaborated with founders on ML-product strategy, translating hiring workflows into data-driven features

The model reviewed more resumes in a day than I probably will in my lifetime.

🤖 Machine Learning Engineer @ IIIT Delhi Hackathon (Finalist) | Aug 2024

• Trained ResNet50 classifier achieving 94% accuracy; reduced training time by 30% via mixed-precision training and optimized data loading

Turns out standing on the shoulders of pre-trained giants is a solid strategy.

💻 ML & Backend Developer @ Smart India Hackathon (Finalist) | Oct 2024

• Built ML-powered hospital queue system reducing predicted patient wait time by 28%
• Deployed Flask APIs on Google Cloud Platform with 99.5% uptime

Built this during the hackathon, survived on coffee and optimism.

Built a structured magnitude pruning pipeline for U-Net achieving 97.3% parameter reduction and 92% FLOP reduction while maintaining IoU > 0.95 on MoNuSeg dataset. Fully reproducible via config.

Tech: Python, TensorFlow, CNN, Model Optimization

97.3% smaller, 100% as accurate. Marie Kondo would be proud.

Implemented a GPT-style Transformer Language Model from scratch in PyTorch — multi-head self-attention, positional encoding, residual connections; modular pipeline with configurable depth and context length.

Tech: Python, PyTorch, Transformers

Built to answer "But how does attention actually work?" Spoiler: it's all matrix multiplication.

Applied LoRA adapters on PEGASUS via HuggingFace PEFT; reduced trainable parameters 767M → 1.57M (99.8%) and achieved 27× faster training vs. full fine-tuning. Served via FastAPI endpoint with Docker packaging.

Tech: Python, PyTorch, HuggingFace, PEFT, LoRA, FastAPI, Docker

Because training 767M parameters is expensive, and I'm not Google.

Python	C++	SQL	PyTorch	TensorFlow	Scikit-learn
NumPy	Pandas	HuggingFace	OpenCV	Flask	GCP
Docker	Git	Jupyter	Linux	VS Code	GitHub

• 🥇 Finalist - Smart India Hackathon 2024
• 🥇 Finalist - IIIT Delhi Hackathon 2024
• 💻 Codeforces - Pupil (1300+ rating)
• 🎯 200+ problems solved across competitive programming platforms

Currently grinding LeetCode and Codeforces. Send encouragement (or better yet, test cases).

_{⚡ Open to Summer 2026 research internships and collaboration opportunities in ML/AI}

_{P.S. If you're reading this, my SEO worked or you're really bored. Either way, hi!}