🧠 GenAI Demos Portfolio

This repository contains a collection of demos showcasing Generative AI techniques.
The notebooks are built with the latest Python libraries, support multiple GPU providers, and include references and explanations where relevant.

👉 The goal: demonstrate practical implementations of GenAI concepts — such as image generation, RAG methodologies, agent orchestration with LangGraph, A2A and MCP, and multi-modal AI — in a way that is both educational and extensible.

---

📂 Contents & Topics

🔹 Tool & Agent sharing across team**s (A2A, MCP, FastAPI)

• a2a_olympics.ipynb → leverages A2A protocol to select optimal agents for selected tasks. Compares results against contract with a general-purpose agent.
• archival_restore.ipynb → applies model context protocol (MCP) to develop an image enhancement approach optimal for OCR exploitation on a 130 year-old diary.

🔹 Multi-Agent Workflows (LangGraph / LangChain)

• mtg-card-generator.ipynb → shows the benefits of using multiple agents for a complex task: building new creature cards for a popular trading card game (Magic: The Gathering)
• langGraph-multi-agent-workflow.ipynb → combines a search agent + image generation agent (OpenAI) to generate a grid of famous athletes by country.
• langGraph-alphabet.ipynb → creates an AI-powered children’s alphabet poster with words + images.

🔹 GANs (Generative Adversarial Networks)

• dcgan_faces_tutorial.ipynb → trains a GAN on a celebrity dataset, demonstrating generator vs discriminator competition and pitfalls of the approach.
• GAN_for_images_MLP_vs_CNN → Proves "NN architecture matters". Compares the results of building a GAN with a dense NN vs a CNN. The dense NN fails as it does not have a sense of space.

🔹 CLIP (Contrastive Language–Image Pretraining)

• CLIP-from-ground-up.ipynb → builds a CLIP model from scratch on Flickr dataset (images + captions).
• CLIP_tuning.ipynb → fine-tunes a CLIP model for domain-specific tasks (example: aerial imagery).

🔹 Retrieval-Augmented Generation (RAG)

• Building_a_multimodal_RAG.ipynb → queries both stored text and images with multimodal input.

🔹 Multimodal orchestration

• Agentic-Multimodal.ipynb → solves a common problem with creating images with text by combining an orchestrator agent, image generation through a CLIP model, and a research tool for factual data. Builds posters of leaders with names & time in office, maps of the world with country & capital with a picture of a nation's celebrity, and maps a great circle route between two places on the globe.

🔹 Diffusion Models

• annotated_diffusion.ipynb → step-by-step theory & math of progressive noise addition with MNIST fashion dataset.
• score_based_diffusion.ipynb → implements score-based diffusion with ODE solver.
• Stable_Diffusion.ipynb → uses HuggingFace diffusion models with CLIP text encoder for text-to-image generation.

🔹 Stability.ai API

• Stability_API_Demo.ipynb → demonstrates masking, overlaying, background replacement, and other image-editing tasks via Stability.ai API.

🔹 Variational Autoencoders (VAE)

• VAE_Example.ipynb → compares a standard autoencoder vs a variational autoencoder to explain the power of VAEs.

---

🚀 Tech Stack

• Python (latest libraries)
• PyTorch / HuggingFace / LangChain / LangGraph
• Diffusion, GANs, CLIP, RAG, VAEs
• APIs: Stability.ai, OpenAI

---

📊 Results

Sample outputs are stored in the results/ folder.

🔎 These are not required to run the demos — they simply illustrate recent experiments.

---

🧩 How to Use

1. Clone the repo:

   git clone https://github.com/your-username/genai-demos.git