AI & LLM Projects – Shubham Aggarwal 👋

Welcome to my AI & LLM project portfolio! I’m passionate about cutting-edge machine learning research and real-world AI applications. This repository showcases projects across Large Language Models (LLMs), data science competitions, and classic ML/NLP tasks, highlighting the impact, tools, and methods behind each.

💡 Explore my achievements, skills, and certifications—updated regularly to reflect my AI journey.

🌐 Check out all my projects on GitHub

---

🔗 Profiles

• GitHub: github.com/Shubham2376G
• LinkedIn: linkedin.com/in/shubham-aggarwal
• Email: shubham_agg@alumni.iitm.ac.in
• Resume: Resume_AI
• Certificates: Here

---

📜 Certifications

• IBM Data Science Professional Certificate (10-course specialization)
• TensorFlow 2.0: Deep Learning and Artificial Intelligence
• Machine Learning Specialization (3 courses)
• Transformers for Natural Language Processing
• Machine Learning: Natural Language Processing in Python (V2)
• Modern Computer Vision GPT, PyTorch, Keras, OpenCV4 in 2024!
• Generative AI, from GANs to CLIP, with Python and Pytorch

---

🔬 LLM Research Study

Knowledge Distillation Implemented a teacher–student learning pipeline to compress large LLMs into smaller ones, reducing inference cost while retaining high performance. Key Insight: A well-designed student model can achieve near-teacher accuracy with far fewer parameters, making deployment more efficient and fast.

Speculative Decoding Implemented speculative decoding to accelerate LLM inference by letting a smaller draft model propose multiple tokens, which are then selectively verified by a larger target model. Key Insight: Parallelizing token generation using a lightweight draft model significantly speeds up decoding while maintaining the accuracy of the main model.

1-bit LLM Reimplemented Microsoft’s 1-BitNet, which leverages 1-bit quantization for training large-scale LLMs, drastically reducing memory and communication cost. Key Insight: 1-bit quantization, combined with error compensation, enables near full-precision accuracy while significantly lowering training overhead.

Chain-of-Thought (CoT) Decoding Developed a prompt-engineering framework that guides LLMs through sequential reasoning steps. Improves accuracy and explainability in reasoning tasks.

LangGraph Chatbot Learning Created Python scripts to learn LangGraph and build a local chatbot with structured memory and human-in-the-loop messaging. Implemented multiple components including ReAct reasoning, parallelization, state management, and long-term memory to understand LangGraph workflows.

Fine-Tuning Qwen-2 Vision-Language Fine-tuned Qwen-2 VL multimodal model using LoRA and LlamaFactory for structured data extraction from product images.

🏆 Competition Work

Data Science Championship 2024 Developed a model to classify patent papers into EPO categories, including text preprocessing with NLTK (stop words removal, stemming, lemmatization) and TF-IDF vectorization.. Used an ensemble of CNN and RNN models to improve accuracy to 68%, ranked Top 10 among 1000+ competitors, and presented the solution to a jury.

Data Vizz Contest 2024 Built an interactive Tableau dashboard to analyze world cuisine trends with filters for country, cuisine type, and ratings. Performed a SWOT analysis on global cuisine patterns to highlight strengths, weaknesses, opportunities, and threats.

AMEX Cricket Analytics Contest 2024 Performed advanced feature engineering on cricket data. Trained stacked ensembles (XGBoost, LightGBM, CatBoost) with Optuna hyperparameter tuning to maximize predictive accuracy.

🤖 Machine Learning & NLP Projects

AI Agents (Healthcare) Created a group chat system using Microsoft AutoGen with specialized AI agents (Fitness & Nutritionist). Built agents to deliver personalized workout and diet recommendations, simulating an interactive health advisory team.

FAQ Generator Built an agentic workflow to automatically generate FAQ sections for websites, including relevant hyperlinks. Leveraged SLMs to analyze website content, extract key topics, and create concise, well-structured FAQ entries.

Reinforcement Learning Workshop Organized and led a workshop on Q-learning and advanced RL methods (OPRO, GRPO).

IBM FalconX Rocket Landing Prediction Prepared the dataset by fetching data via RESTful APIs and scraping sources like Wikipedia using BeautifulSoup. Explored and visualized data with Folium and Seaborn heatmaps, and trained an SVM model achieving 88.8% accuracy.

Sentiment Analysis Implemented a text classifier for positive/negative/neutral sentiment categorization.

Email Spam Detection Trained a Naive Bayes spam classifier with engineered features (keywords, sender info, formatting patterns).

📚 Research Engagement

In addition to projects and competitions, I also contribute to research.
I maintain a Preprints folder containing my own works:
👉 My Research Preprints

These reflect my initial explorations into AI research directions and demonstrate my ability to translate ideas into written scholarly form.

Alongside this, I have studied and annotated 25+ research papers to deepen my theoretical understanding of AI. These span:

• Scaling & Efficiency → Scaling Laws, Chinchilla, Broken Scaling Laws, BitNet, Test-Time-Training
• Reasoning in LLMs → Chain-of-Thought Prompting, CoT Decoding, Logic of Thought, Scheming LLMs
• Model Compression & Optimization → Knowledge Distillation, Batch Normalization, Differential Methods, Ramanujan’s Randomly Weighted Networks
• Generative Models → Diffusion vs. Autoregressive Models, Qwen3 Technical Report
• Cross-disciplinary Methods → Betti Numbers in Topology, AI for Data Analysis

A dedicated folder with my notes and summaries is available here:
👉 Research Papers – Literature Notes

Reading and annotating research papers not only strengthens my theoretical foundation but also sparks creative new ideas for projects, experiments, and applications.