Agentic AI & GenAI Playground

This repository is a small, curated set of projects I’ve built in my spare time to explore how to use modern Agentic AI and GenAI tools, concepts and frameworks in practical, real-world settings.

It focuses on:

• Using LLM-driven agents to design, run, and evaluate end-to-end ML workflows.
• Embedding GenAI models into real data/ML stacks (classic ML, computer vision, video) instead of one-off demos.
• Optimising prompts and agent programs (e.g. via DSPy/MIPROv2-style) rather than treating prompts as static text.
• Building evaluation loops for agents and models (end-to-end vs component-based evaluations, LLM-as-a-judge).
• Exploring current AI engineering tools and patterns in a fast-moving ecosystem (agent frameworks, multimodal LLMs, and GenAI video).

Each folder is a self-contained project with its own README.md, code, and setup, designed to be read like a real-world pattern rather than a toy example.

---

Projects

1. AgenticClassicML

Folder: AgenticClassicML/
Readme: AgenticClassicML/README.md

What it is

An example that uses Plexe (built on top of smolagents) to automatically build classical ML baselines.

Instead of manually writing the entire pipeline, the project:

• Uses a LLM-powered multi-agent to:
  • Inspect a real tabular dataset,
  • Generate training and evaluation code in Python,
  • Train simple models (logistic regression, decision trees),
  • Package the pipeline and artifacts into a reusable tarball.
• Constrains the solution to simple, interpretable models to create strong, transparent baselines.
• Demonstrates an LLM-driven AutoML-style workflow:
  • Data loading and splitting,
  • Agent-driven model building,
  • Prediction and inspection,
  • Artifact creation and reuse.

Why it matters

This project is designed to signal:

• Hands-on experience using agentic frameworks (Plexe + smolagents) rather than just talking about them.
• A solid foundation in classic ML (logistic regression, decision trees, tabular baselines).
• The ability to design and run LLM-orchestrated workflows end-to-end (data prep → training → evaluation → packaging).

For details, setup instructions, and code walkthrough, see the dedicated readme:
➡️ AgenticClassicML/README.md

---

2. AgenticCV

Folder: AgenticCV/
Readme: AgenticCV/README.md

What it is

An example that uses DSPy to build an agentic computer vision classifier on the MVTec AD capsule dataset.

Instead of training a conventional CNN, the project:

• Wraps a multimodal LLM (e.g. gpt-4o) in a Chain-of-Thought DSPy module to:
  • Inspect capsule images and reason about visual defects,
  • Classify each image into crack, scratch, poke, faulty_imprint, squeeze, or good.
• Uses MIPROv2 to automatically optimize the prompt/program:
  • Baseline Chain-of-Thought model: ~60% exact-match accuracy,
  • MIPROv2-optimized program: ~76% exact-match accuracy.
• Exports the optimized programs as JSON files that capture:
  • The final prompt,
  • The DSPy configuration,
  • The structure of the compiled model.

Why it matters

This project is designed to signal:

• Practical experience using DSPy for agentic CV, not just calling a vision API once.
• Ability to turn a folder of images into a structured evaluation pipeline using dspy.Example, stratified splits, and exact-match metrics.
• Familiarity with program-level optimization (MIPROv2) to systematically improve LLM-based models and save the resulting configurations for reuse and deployment.

For details, setup instructions, and code walkthrough, see the dedicated readme:
➡️ AgenticCV/README.md

3. GenAIVideoGenerationVeo3

Folder: GenAIVideoGenerationVeo3/
Readme: GenAIVideoGenerationVeo3/README.md

What it is

An example that uses Gemini 2.5 + Veo 3 to automatically generate multi-scene videos (e.g. vlogs) from a single high-level idea.

Instead of manually editing video or hand-crafting every shot, the project:

• Uses Gemini 2.5 Pro to:
  • Expand an idea (e.g. “tourist kangaroo visiting Paris”) into a sequence of structured scene prompts,
  • Return scenes as JSON (Scene / SceneResponse) with both positive and negative descriptions.
• Uses Veo 3 to:
  • Generate a short video clip per scene with cinematic 16:9 framing,
  • Respect style constraints (e.g. realistic, 4K, vlog style).
• Uses MoviePy to:
  • Stitch all scene clips together,
  • Produce a final vlog.mp4 ready to share.

Why it matters

This project is designed to signal:

• Practical experience orchestrating multi-step GenAI video workflows (text → scenes → clips → merged video).
• Ability to combine LLM prompt engineering, structured JSON outputs, and video generation APIs in a single pipeline.
• Familiarity with real-world tooling such as google-genai and moviepy to build portfolio-ready GenAI demos (e.g. automated vlogs, short stories, or marketing clips).

For details, setup instructions, and code walkthrough, see the dedicated readme:
➡️ GenAIVideoGenerationVeo3/README.md

Structure

As more projects are added, this repo will group them by theme, for example:

• AgenticClassicML/ – Agent-driven classical ML baselines (Plexe, smolagents, scikit-learn).
• AgenticXXXX/ – Agentic workflows for other modalities (e.g. image, text, graphs, retrieval, etc.).

Each project is:

• Self-contained – its own environment and instructions.
• Reproducible – clear setup and run steps.
• Explainable – focused readme and rationale for design choices.

---

How to Use This Repo

• Browse the project folders to find the scenario you care about.
• Start with AgenticClassicML if you want a concrete, tabular ML example using Plexe.
• Use these projects as:
  • Portfolio pieces,
  • Starting points for your own agentic workflows,
  • Conversation drivers with teams building Agentic AI / GenAI products.

More projects will be added over time, covering different data types, tools, and deployment patterns.