CastIt — The ML Casting Director

Inspiration

Choosing the right machine learning model for a task is a lot like casting a movie. there are thousands of candidates, each with different strengths, and picking the wrong one can sink the whole production. We wanted an agent that acts like a casting director: reviewing every "audition" (experiment), checking the numbers, and confidently recommending who gets the part.

What it does

CastIt is an AI agent built on Google's Agent Development Kit (ADK) that analyzes a dataset of 1,000 real ML experiments spanning 10 OpenML classification tasks. Ask it things like:

• "Who's the best model for this role?" → returns the top-performing model, backed by its actual accuracy score
• "Give me the shortlist" → returns the top 5 candidates for a task
• "Compare these two stars" → head-to-head performance comparison between two model families
• "Give me the full audition report" → a complete performance summary for a task

Every answer is grounded in real data- no invented numbers, no guessing.

How we built it

• Python analysis layer (pandas) that loads and queries the cleaned OpenML experiment dataset
• Google ADK agent with four tools (get_best_model, get_top_models, compare_models, get_task_summary) that the LLM calls based on the user's question
• Gemini as the underlying model, reasoning over tool outputs to give clear, conversational answers

Challenges we ran into

Getting the ADK agent to run reliably in both the CLI and the local web UI took real debugging, from Python package-naming rules (no hyphens allowed) to making sure file paths worked no matter which directory the agent was launched from. Small config details mattered a lot more than expected for a "casting director" to actually take the stage.

Accomplishments we're proud of

A working end-to-end agent that reliably answers real analytical questions over a 1,000-row dataset with grounded, tool-verified answers , running cleanly in both the terminal and ADK's web interface.

What's next

• Expand beyond the current 10 tasks
• Add a "director's commentary" mode that explains why a model won, not just that it did
• A simple web front-end styled like a casting call board

Built With

• gemini
• google-adk
• openml
• pandas
• python