Welcome to TheGradientPath 🚀

Your comprehensive learning journey through modern Machine Learning, Deep Learning, and Artificial Intelligence - from fundamentals to production systems.

This repository is a complete educational resource that bridges theory with practice, covering everything from foundational neural networks to cutting-edge AI agent systems and production ML deployment. Each project is designed to be hands-on, practical, and production-ready, with clear documentation, video tutorials, and runnable code.

---

📚 Table of Contents

• AI Agents
• Deep Learning with Keras
• PyTorch Projects
• LLM Fine-Tuning
• RAG Systems
• Real-World Production Projects
• MCP Protocol
• AI Security Engineering
• Getting Started

---

AI Agents

Comprehensive AI Agent Framework Benchmark

Location: AiAgents/AgentFrameworkBenchmark/

A production-grade comparison of 7 major AI agent frameworks implementing identical multi-agent systems to provide objective, real-world benchmarks.

Video Tutorial

Watch on YouTube - Complete framework comparison and implementation guide

Frameworks Compared

1. LangChain/LangGraph (🥇 284/360) - Best overall, maximum flexibility
2. OpenAI Agents (🥈 277/360) - Minimal code, native MCP support
3. CrewAI (🥉 249/360) - Rapid prototyping, simple delegation
4. LlamaIndex (227/360) - Balanced workflow architecture
5. AutoGen (195/360) - Enterprise async infrastructure
6. Semantic Kernel (178/360) - Microsoft ecosystem integration
7. Vanilla Python - Baseline with zero framework overhead

What's Benchmarked

• ✅ Agent Orchestration - Multi-agent coordination and routing
• ✅ Tool Integration - Custom tool creation and execution
• ✅ State Management - Complex state handling across agents
• ✅ Memory Management - Persistent conversation history
• ✅ MCP Server Integration - Model Context Protocol support
• ✅ Production Features - Guardrails, token tracking, structured output

System Architecture

Each implementation includes:

• Orchestrator Agent - Routes queries to specialized agents
• Legal Expert Agent - Handles law and legal topics
• Operational Agent - Manages programming and general queries
• Tools - Weather API, calculator, web search
• MCP Integration - Extended capabilities via Model Context Protocol

Full Documentation

---

Deep Learning with Keras

Modern Neural Network Implementations

Location: Keras/

Production-ready deep learning implementations using TensorFlow and Keras, from fundamentals to advanced architectures.

1. Image Classification with MLP

Path: Keras/ImageClassificationWithMLP/

• Dataset: MNIST handwritten digits
• Architecture: Multi-layer perceptron with Dropout and BatchNormalization
• Features: TensorBoard integration, Visualkeras architecture diagrams
• Tools: Dense layers, functional API, comprehensive logging

2. Transformer-Based Text Generation

Path: Keras/transformers/text_generation/

• Task: Natural language generation from scratch
• Architecture: Complete Transformer implementation
• Components: Multi-head self-attention, positional encoding, feed-forward networks
• Features: Custom training loop, text preprocessing, generation sampling

3. Text Generation with KV Cache

Path: Keras/transformers/kv_cache_for text_gen/

• Optimization: Key-Value cache for efficient inference
• Performance: Dramatically reduced computation during generation
• Architecture: Modified Transformer with caching mechanism
• Use Case: Production LLM inference optimization

4. Time Series Forecasting with Transformers

Path: Keras/transformers/time_series_forecast/

• Task: Stock price prediction using Transformers
• Data: Synthetic financial time series
• Architecture: Transformer adapted for sequential prediction
• Features: Temporal embeddings, MinMax scaling, visualization

Key Learning Points:

• Building Transformers from scratch in Keras
• Multi-head attention mechanisms
• Positional encoding strategies
• KV cache optimization techniques
• Adapting Transformers for different domains

---

PyTorch Projects

CNN Image Classification

Location: Pytotch/CnnImageClassification/

Full README

Fashion-MNIST CNN Classifier

• Dataset: 70,000 images of 10 clothing categories
• Architecture: 2-layer CNN with BatchNorm
  • Conv2d(1→16) + BatchNorm + ReLU + MaxPool
  • Conv2d(16→32) + BatchNorm + ReLU + MaxPool
  • Fully Connected (512→10)
• Performance: ~85-90% validation accuracy
• Features:
  • Automatic dataset download
  • GPU acceleration support
  • Model checkpointing
  • Training visualization
  • Real-time progress monitoring

Key Learning Points:

• Convolutional neural networks fundamentals
• Batch normalization for training stability
• PyTorch DataLoader and Dataset classes
• Model training and evaluation pipelines

---

LLM Fine-Tuning

Advanced Parameter-Efficient Fine-Tuning Techniques

Location: LLMFineTuning/

State-of-the-art techniques for efficiently fine-tuning large language models for specific tasks.

1. All PEFT Techniques From Scratch

Path: LLMFineTuning/all_peft_tecniques_from_scratch/

Complete implementation of Parameter-Efficient Fine-Tuning methods:

• LoRA (Low-Rank Adaptation) - Inject trainable low-rank matrices
• Prefix Tuning - Learn soft prompts prepended to inputs
• Adapter Layers - Small bottleneck layers inserted into models
• IA³ (Infused Adapter by Inhibiting and Amplifying Inner Activations)

Why PEFT?

• Train only 0.1-1% of model parameters
• Reduce memory requirements by 90%
• Maintain performance close to full fine-tuning
• Enable multi-task learning with parameter isolation

2. GRPO Reasoning with Unsloth

Path: LLMFineTuning/GRPO_REASONING_UNSLOTH/

Advanced reasoning capabilities through Gradient-based Reward Policy Optimization:

• Model: Google Gemma 3 1B with 4-bit quantization
• Technique: GRPO (combines PPO benefits with gradient-based optimization)
• Task: Mathematical reasoning with structured outputs
• Features:
  • LoRA rank-32 adaptation
  • 4-bit quantization for memory efficiency
  • vLLM acceleration for fast inference
  • Structured reasoning format (<reasoning> and <answer> tags)

Performance Gains:

• Models learn to show reasoning steps
• Improved accuracy on complex problems
• Better interpretability of model decisions

3. Supervised Fine-Tuning with Tool Choice

Path: LLMFineTuning/SFT_HF_TOOL_CHOICE/

Teaching models to intelligently select tools:

• Model: HuggingFace SmolLM2-135M
• Task: Tool selection based on user queries
• Dataset: 10,000 synthetic examples with tool annotations
• Technique: Supervised Fine-Tuning with custom special tokens
• Use Case: Building function-calling capabilities in smaller models

Real-World Application:

• Enable LLMs to use external tools (calculators, APIs, databases)
• Reduce reliance on large models for specialized tasks
• Build cost-effective AI assistants

---

RAG Systems (Retrieval-Augmented Generation)

Advanced RAG Architectures

Location: Rag/

Production-ready Retrieval-Augmented Generation systems that enhance LLM responses with external knowledge.

1. Dartboard RAG

Path: Rag/dartboard/

Full README

Balanced Relevance and Diversity Retrieval

Based on the paper: "Better RAG using Relevant Information Gain"

Key Innovation:

• Problem: Standard top-k retrieval returns redundant documents
• Solution: Optimize combined relevance-diversity score
• Result: Non-redundant, comprehensive context for LLMs

Features:

• Configurable relevance/diversity weights
• Production-ready modular design
• FAISS vector store integration
• Oversampling for better candidate selection

Algorithm:

combined_score = diversity_weight * diversity + relevance_weight * relevance

When to Use:

• Dense knowledge bases with overlapping information
• Queries requiring diverse perspectives
• Avoiding echo chambers in retrieval

2. Hybrid Multivector Knowledge Graph RAG

Path: Rag/hybrid_multivector_knowledge_graph_rag/

Full README

The Most Advanced RAG System - 11+ Graph Traversal Algorithms

Revolutionary Features:

• Knowledge Graph Engineering with Neo4j
• Multi-Vector Embeddings for nuanced retrieval
• 11+ Graph Traversal Algorithms:
  • K-hop Limited BFS
  • Depth-Limited DFS
  • A* Search with heuristics
  • Beam Search
  • Uniform Cost Search (UCS)
  • Context-to-Cypher query generation
  • LLM-powered intelligent filtering

Architecture:

1. Vector Retrieval - Initial similarity search
2. Graph Traversal - Navigate knowledge relationships
3. Entity Extraction - LLM-powered entity identification
4. Dynamic Querying - Context-aware Cypher generation
5. Intelligent Ranking - Multi-factor relevance scoring

Why Knowledge Graphs?

• Discover hidden connections across concepts
• Follow chains of reasoning
• Understand complex relationships
• Navigate multi-hop queries intelligently

Use Cases:

• Research and academic knowledge bases
• Legal document analysis
• Scientific literature review
• Complex domain expertise systems

3. Vision RAG

Path: Rag/vision_rag/

**Multimo