AI
Conversational Querying at the NFL Combine
Scouts and coaches from all 32 NFL teams use Surface tablets running the NFL Combine app to review test results and player data. This year, Southworks added an AI assistant that answers natural-language questions against live combine data. Amiani asked “whiteouts over 6 feet,” got player names and heights back, then asked for a vertical jump comparison across those three players. A follow-up query found Maxwell Hearstone as the 2025 fastest 40-yard dash runner and placed him against the top performers from the past three years.
“So he didn’t make it but was really close.” — Pablo Amiani
The Azure Stack Behind the Assistant
The backend runs on Azure Cosmos DB storing JSON documents for each combine test, GPT-4o1 deployed through Azure AI Foundry, and a Semantic Kernel agent connecting both. Before writing any code, Amiani pasted a Cosmos JSON document into the Azure AI Foundry playground as the system prompt, asked “top five performers in the 40 dash,” and confirmed the model could identify player names, positions, jersey numbers, and test results from raw JSON. That sanity check took under two minutes.
“This looks really good to me.” — Pablo Amiani
GitHub Copilot Agent Mode Builds the Semantic Kernel Agent
GitHub Copilot agent mode scaffolded the entire Semantic Kernel agent in one pass. Amiani gave it a Semantic Kernel reference snippet from GitHub, a sample JSON file matching the Cosmos DB structure, and a tasks.txt listing the query patterns needed, for example “who ran the 40 dash in less than 4.5 seconds.” The generated agent.py included four plugin functions backed by Cosmos DB queries, a system prompt derived from the sample JSON, and the Azure OpenAI chat completion service pointed at GPT-4o1. The console app live query (09:22) returned two players with sub-4.3 times on first run.
From Console to Web UI in One Copilot Edit
Copilot edit mode rewrote agent.py in a second pass. One instruction: “replace the console with a chatbot UI with Gradio.” Copilot scanned every line, showed a diff removing the console and adding the Gradio library, and the updated app launched a web interface with a chatbot component and text input. The Gradio web UI live query (11:32) returned a list of players under 4.4 seconds in natural language. Two Copilot prompts, a working web AI assistant.
“you are not using these tools, you are missing out” — Pablo Amiani
Notable Quotes
you are not using these tools, you are missing out Pablo Amiani · ▶ 12:04
This looks really good to me. Pablo Amiani · ▶ 8:50
So he didn’t make it but was really close. Pablo Amiani · ▶ 2:57
Key Takeaways
- GitHub Copilot agent mode generated a working Semantic Kernel agent with Cosmos DB queries from a three-file spec in minutes.
- GPT-4o1 parsed raw Cosmos DB JSON and answered combine performance questions without any schema translation layer.
- A single Copilot edit-mode prompt swapped a console app for a Gradio web interface without breaking the agent.