Vector-Based Search (FAISS): Used FAISS to generate and store embeddings, enabling efficient document retrieval.
Text Chunking & Ranking: Optimized RAG pipelines for accurate retrieval, experimenting with chunk sizes, metadata, and hybrid ranking (embedding similarity + BM25).
Guardrails & Moderation: Developed safeguards for LLM-generated responses (e.g., hallucination detection, toxicity checks).
Syntax & Control Flow: Mastered Python fundamentals, including variables, data types (int, float, str, list, dict), loops (for, while), and conditionals (if/else).
Functions & Modules: Built reusable code with functions, imported built-in and custom modules (math, random, os), and implemented error handling (try/except).
File Handling & Data Processing: Learned to read/write text and CSV files, process structured data using Pandas, and perform statistical computations using NumPy.
Hallucination Reduction: Used retrieval-augmented generation (RAG) and reinforcement learning to minimize incorrect LLM outputs.
API Interactions: Used the requests library to fetch and parse JSON data from public APIs.
Project 1: Developed a complete data processing pipeline handling file input, computation, and structured output.
Project 2: Built a prompt-based Q&A bot using a local LLM and LangChain for structured retrieval.
Project 3: Created a RAG-powered chatbot that retrieves context-aware answers from a user-defined corpus.
2️⃣ Module 2 ⏳ ~2 months 🕒 3-hour sessions 📅 2 times per week
Training and Fine-Tuning Generative AI Models. Deployment, Testing, Cloud usage of Generative AI and LLMs
Parameter-Efficient Fine-Tuning (LoRA): Learned LoRA (Low-Rank Adaptation) to fine-tune LLMs with minimal computational overhead.
Feature Extraction: Extracted embeddings from pre-trained models for downstream tasks like classification.
Local Model Deployment: Hosted fine-tuned models locally via OpenAPI and tested inference endpoints.
Generative Adversarial Networks (GANs): Studied adversarial training dynamics and strategies to prevent mode collapse.
Transformer Architecture: Learned Transformer architecture, including self-attention mechanisms, multi-head attention, and positional encodings, crucial for LLM effectiveness.
Project: Delivered a fine-tuned LLM-powered AI system, demonstrating an optimized pipeline from training to deployment.
3️⃣ Module 3 ⏳ ~3 months 🕒 3-hour sessions 📅 2 times per week
Deep Learning Foundations with PyTorch for Researchers
Machine Learning Fundamentals: Studied bias-variance tradeoff, overfitting, bagging vs. boosting, gradient descent, and regularization techniques (L1/L2).
Tensors & Autograd: Learned PyTorch tensors, performed matrix operations, and used automatic differentiation for gradient computation.
Deep Learning & Neural Networks: Implemented neural networks from scratch, explored activation functions, dropout, batch normalization, and transfer learning.
Computer Vision: Worked with convolutional neural networks (CNNs), pooling layers, and object detection models like YOLO and Faster R-CNN.
Coding & Algorithms: Practiced Python coding for ML tasks, bigram extraction, binary tree traversal, and graph algorithms.
Data Pipelines: Worked with DataLoaders, mini-batches, and epochs to train models efficiently.
Custom Model Design: Built a transformer-inspired model, introducing attention mechanisms.
Practical Problem-Solving: Developed machine learning solutions for fraud detection, customer churn prediction, supply chain optimization, and anomaly detection.
Cloud Deployment: Deployed models on AWS SageMaker
Project: Built an end-to-end AI pipeline with model deployment, monitoring, and optimization for real-world AI applications.