Shubham Pundhir

About Shubham Pundhir

I’m Shubham Pundhir, a passionate Machine Learning Engineer and Researcher currently pursuing my M.Tech (ECE, ML specialization) at Delhi. Over the years, I’ve worked on many hands-on projects—bridging cutting-edge research in Computer Vision and NLP with real-world, production-ready solutions. Here’s a snapshot of what I bring to the table: 1. Research & Vision Systems: a). e-MARG (Infosys CAI): PoC for an automatic video-based road inspection tool commissioned by India’s Ministry of Rural Development, MoRD. b). City Surveillance & OCR: Engineered a YOLOv5-based detector to locate bus display panels in CCTV feeds. SDE & Simulation: a). EMF-3D (PowerGrid/EDC Lab): Development of a real-time 3D electromagnetic-field visualization tool (Python + SQL backend, PHP/CSS/HTML interface). Natural Language Processing & LLMs: a). Built From Scratch Tokenizers & Embeddings: b). Recipe Generation (LSTM → GPT-2)- Preprocessed five cuisine datasets (51 k+ recipes), fine-tuned GPT-2 so raw ingredient lists translate into full recipes (quantities, instructions, titles). Integrated RAG to augment system prompt to context-aware prompts. c). Claim Normalization & Check-Worthiness. d). Gendered Abuse Detection (Indic Languages): Fine-tuned IndicBERT, MuRIL, and Abuse-XLMR on multilingual social data—improving cross-lingual transfer for hate-speech classification. Data Science & Predictive Modeling- Built scikit-learn pipelines to predict MPCE (Monthly Per Capita Expenditure) using HCES (Household Expenditure Consumption Survey) data. Technical Tools and Skillsets: i. Languages: Python, C++, SQL, MATLAB ii. Frameworks/Libraries: PyTorch, TensorFlow/Keras, Hugging Face Transformers, scikit-learn, OpenCV, WandB, Docker, Git, Linux/Ubuntu Architectures: CNN, YOLO, U-Net, Vision Transformers (ViT), BERT, LSTM, GPT-2, BART, T5 iv. Data Science: EDA, SVMs, Decision Trees, PCA, TF-IDF, Cosine Similarity, XG-Boost, KNN.** Why I Love What I Do!** I thrive on the blend of research rigor and practical impact. Whether I’m fine-tuning a segmentation model to flag potholes or crafting a GPT-2 prompt that writes delicious recipes, I’m fueled by curiosity and the joy of building. Outside code, you’ll find me blogging about gratitude and exploring the latest AI papers—always looking to learn, collaborate, and push boundaries. Let’s connect if you’re excited about AI-driven solutions in GovTech, finance, healthcare, social good, or consumer applications. I’m always up for a meaningful chat (and a strong cup of coffee).

Experience

1. Graduate Research Intern

   Complex Systems Laboratory

   Jan 2025 — Present · New Delhi, IN

   Project title:\"RecipeGPT2-Your-Own AI-Chef\"- Designed and implemented an end-to-end novel recipe generation pipeline by fine-tuning GPT-2 on culinary data(RecipeDB)- Scraped and curated a dataset from RecipeDB, then performed comprehensive data cleaning and JSON-structuring- Trained a Byte Pair Encoding (BPE) tokenizer vocab size) with custom special tokens (, etc.)- Fine-tuned GPT-2 using structured input–output formatting. Used H100 GPU with mixed precision training to utilize resources judiciouly- Leveraged fine-tuning methods (LoRA and QLoRA) to reduce memory dependency and accelerate training- Conducted extensive feature engineering and prompt-design experiments to enhance generation quality(Chain of Thoughts)- Evaluated model outputs via BLEU scoring, BERTF1 score for semantics- Built an inference module to generate full recipes (title, ingredients with quantities, instructions) from raw ingredient lists- Presented methodology and results to my guide Dr. Ganesh Bagler, as a part of my capstone project.

Education

• Indraprastha Institute of Information Technology, Delhi

  Master of Technology - MTech

  2024 — 2026

• Punjab Technical University

  Bachelor of Technology - BTech

  2017 — 2021

• St Fidelis School,Aligarh

  HSC

  2015 — 2017

• Indian Institute of Technology Jodhpur

  Master of Technology - MTech

  2024 — 2024