🧠 LLM Playbook with Experiments

A practical playground for exploring Large Language Models (LLMs), LangGraph, and Deep Learning fundamentals — from linear algebra all the way to multi-agent orchestration and long-running durable systems.

This repository represents my continuous journey as a Machine Learning Engineer and Software Engineer experimenting with the internals of AI systems and production-ready agent architectures.

---

🚀 Project Overview

This project bridges core ML theory with applied LLM engineering.
Each folder focuses on a distinct research or engineering concept:

Folder	Focus Area	Key Topics
deep_learning/core_foundation/	Mathematical fundamentals	Linear Algebra, Matrix operations, Backprop basics
llms/1_core_foundation_single_agent/	Building foundational LLM workflows	Attention mechanism, finite-state thinking, GPU checks
llms/2_memory_retrieval_integration/	Memory-augmented systems	RAG (Retrieval Augmented Generation), Chroma vector store, Ollama local LLM integration
llms/3_multi-agent-pipeline/	Multi-agent orchestration	LangGraph + LangChain, cooperative agents, research + summarization + answer pipeline
llms/4_long_running_process/	Durable, reliable agents	Queue management, retries, persistence, HITL (Human-in-the-Loop) execution
llms/5_all_about_langgraph/	LangGraph deep dive	Graph nodes, conditional edges, checkpointing, message passing patterns

---

🧩 Key Experiments & Learnings

1️⃣ Core Foundations

• Implemented attention mechanism from scratch to understand transformer internals.
• Built simple finite-state machines to simulate agent reasoning.
• Practiced LangGraph basics (nodes, edges, conditions) to design structured reasoning flows.

2️⃣ Retrieval-Augmented Memory

• Integrated Ollama (local LLMs like Llama-2 / Mistral) with ChromaDB for semantic retrieval.
• Implemented a RAG pipeline to enrich context dynamically during LLM responses.

3️⃣ Multi-Agent Collaboration

• Created Researcher → Summarizer → Answerer workflow.
• Introduced caching, embedding persistence, and modular agents.
• Showcased graph orchestration using LangGraph to coordinate autonomous agents.

4️⃣ Long-Running & Durable Systems

• Implemented durable execution with diskcache and tenacity for retries.
• Added Human-in-the-Loop (HITL) corrections to simulate real feedback and preference learning.
• Designed persistent task states for resumable, production-grade pipelines.

5️⃣ LangGraph Internals

• Studied graph anatomy: nodes vs edges, conditional edges, state channels, checkpointing, and message passing.
• Visualized graphs using Mermaid and Graphviz to understand data flow.

---

⚙️ Tech Stack

Category	Tools / Libraries
LLMs & Agents	LangGraph, LangChain, Ollama
Retrieval	ChromaDB
Persistence / Retry	diskcache, tenacity
Visualization	IPython.display, Graphviz, Mermaid
Core ML	Python, NumPy
Experiment Management	Jupyter Notebooks

---

🌱 Why I’m Building This

Modern AI systems (like ChatGPT, Claude, or Gemini) are not single models —
they are networks of reasoning agents, memory layers, and durable feedback loops.

This repo is my effort to:

1. Understand how LLM agents can be orchestrated like microservices.
2. Explore long-running, stateful AI systems beyond single-prompt inference.
3. Combine theory (math & DL) with practical agent engineering.
4. Prepare for production-level AI development (queue management, retries, persistence, feedback learning).

---

🔭 Future Scope

• Integrate Celery + Redis for distributed task orchestration.
• Add async workflows and pa