At Datathon, we like to give people a breadth of challenges to solve! You're welcome to pick one to solve or all!

🚨IMPORTANT FORENOTE: These are NOT all the challenges we have, just the ones that would be submitted using Devpost! The other challenges as well as a LARGER description for each challenge can be found on our website!

https://www.tamudatathon.com/challenges

Case Closed!:

The grid is your crime scene. Every move leaves evidence. Every turn could be your last. Welcome to Case Closed, a high-stakes AI programming challenge where you’ll train intelligent Python agents to outthink, outmaneuver, and outlast their opponents in a fast-paced grid arena. 

Every tick is a clue. Every direction, a decision. Move too fast, and you’ll collide with your own mistakes; move too slow, and your opponent will box you in. 

Inspired by the classic Tron Lightcycle game, your agents must navigate a 2D grid where every move leaves behind a permanent barrier. One wrong turn — and its case closed.

Will your code be the perfect detective — reading patterns, predicting danger, and solving the puzzle before it’s too late?

Or will it become just another unsolved mystery in the grid?

Objective:
Your mission is to design and train an autonomous agent that can survive longer than its opponent in a series of head-to-head matches. 

Each tick, both agents simultaneously choose a direction (up, down, left, or right). Every cell they move through becomes a wall that cannot be crossed again—by either player. The round ends when one or both agents collide with a wall or each other.

Gameplay Details:

• The arena is a 18 x 20 grid
  
  

• Each agent’s move leaves a trail that acts as a barrier.
  
  

• Agents receive state information each tick (positions, remaining open spaces, opponent’s direction).
  
  

• Colliding with any trail means instant elimination.
  
  

• If both agents crash simultaneously, it’s a draw.

Each tick, your agent receives a JSON payload with key details like:

• Current position (your coordinates and your opponent’s)
  
  

• Board layout (walls, trails, open spaces)
  
  

• Available boosts and remaining turns

• Opponent’s last direction

Build-a-MCP:

Your mission is to create a server that intelligently fetches, filters, and formats data from real-world services (Google Calendar, Notion, AWS, GitHub, etc.) and injects it as context for an AI. The goal is to make the AI assistant vastly more personal, aware, and useful.

The best submissions will be judged not just on technical implementation but on creativity and real-world utility. 

Mai Shan Yun Inventory Intelligence:

The Mai Shan Yun Inventory Intelligence Challenge invites you to design and build an interactive dashboard that transforms raw restaurant data into actionable intelligence.

You’ll be working with a real dataset from Mai Shan Yun including monthly purchase logs, ingredient usage, shipment details, and menu item sales trends.
Your mission? Help the restaurant optimize its inventory: minimizing waste, avoiding shortages, and predicting what to restock and when.

The best submissions will blend data analytics, visual storytelling, and real-world impact to make restaurant management smoother than a perfect stir-fry.

Detective Databricks Mini Challenge (a supporting challenge):

At the heart of every mystery is a trail of clues buried in massive amounts of evidence. Databricks is your high-powered data lab, a crucial platform that unifies Data Warehousing and Data Lake capabilities to help digital detectives process, manage, and analyze petabytes of information using tools like Apache Spark.

Your Mission: This mini-challenge asks you to prove your skills as a data detective. Show us how YOU used the Databricks platform, its notebooks, processing power, or Lakehouse architecture, to manage, clean, analyze, or model the data required to crack your case in one of the other challenges. The best submissions will clearly narrate the data mystery you solved and the specific Databricks features that acted as your crucial magnifying glass.

AI-Powered Regulatory Document Classifier — Datathon Submission

Description

Build an AI-powered assistant that dynamically analyzes multi-page, multi-modal documents to classify them into Public, Confidential, Highly Sensitive, or Unsafe categories. The solution should leverage a configurable prompt library to generate dynamic prompt trees, incorporate Human-in-the-Loop (HITL) feedback to improve accuracy over time, and support both interactive and batch processing with real-time status updates. Pre-processing checks (e.g., legibility, page/image counts) are required. Outputs must include citation-based evidence pointing to the exact pages or images that led to each classification. Optionally, teams may employ two LLMs to cross-verify classifications to reduce HITL needs. The UI should be business-friendly with visualizations, detailed classification reports, audit trails, and easy file management, while ensuring compliance with data privacy and security policies.

Categories

• Sensitive/Highly Sensitive: Content that includes PII like SSNs, account/credit card numbers, and proprietary schematics (e.g., defense or next‑gen product designs of military equipment).  

• Confidential: Internal communications and business documents, customer details (names, addresses), and non-public operational content.

• Public: Marketing materials, product brochures, public website content, generic images.

• Unsafe Content: In any case all content should be evaluated for Child Safety and should not include Hate speech, exploitative, violent, criminal, political news or cyber-threat content.

Evaluation Criteria

• Multi-modal input: accept text, images, and optional video.

• Interactive and batch processing modes with real-time status updates.

• Pre-processing checks: document legibility, page and image count.

• Dynamic prompt tree generation from a configurable prompt library.

• Citation-based results: reference exact pages or images for audit and compliance.

• Safety monitoring: automatically detect Unsafe content and flag for human review.

• HITL feedback loop: enable SMEs to validate outputs and refine prompt logic.

• Double-layered AI validation (optional): two LLMs to cross-verify classifications.

• Rich UI: clear visualizations, detailed classification reports, audit trails, and file management.

• Criteria
  All criteria is listed on our website per challenge! https://www.tamudatathon.com/challenges