Teamfight Tactics's Tutor for newbie players.

Inspiration

When I first started playing Teamfight Tactics (TFT), I found the learning curve incredibly steep. To truly enjoy and master the game, you need to absorb a massive amount of dynamic knowledge—from item recipes to complex unit synergies. I realized that this complexity scares away many potential players. I wanted to build a bridge that allows anyone to jump straight into the fun of strategy without struggling to understand the overwhelming mechanics.

What it does

It is not just a chatbot it is an intelligent AI Tactical Advisor. Powered by Gemini and the advanced Agentic frameworks LangGraph and LangChain, it analyzes the user's specific situation to provide real-time advice. It can answer complex queries about the best items for specific units, calculate optimal trait synchronization, and guide the user toward the strongest board state based on the current "meta."

How we built it

Data Pipeline: We engineered a pipeline using the Riot Games API to fetch match data from Challenger-ranked players on the NA server. We analyze these high-level matches to understand "winning blueprints"—what items they build, what units they prioritize, and how they pivot.
Agentic Workflow: We wrapped this data into custom tools for our Agent. Using LangGraph, we created an orchestrated workflow. The LLM first checks our internal analyzing engine. If the specific data isn't there, it autonomously utilizes web-scraping tools (targeting sites like MetaTFT) or performs a Google Search to ensure the advice is accurate and up-to-date.

Challenges we ran into

Data Normalization: The Riot API often returns legacy object keys (old item names) that don't match the current set. We had to build a robust mapping system to translate this raw data into the up-to-date terminology used by players.
The Vision Constraint: Our original ambition was a Real-time Vision Agent that could "watch" the screen and coach the player live. However, we discovered that without extensive fine-tuning, current Multimodal models struggle to accurately read the specific UI details of the game interface. Since we prioritized accuracy over novelty, we pivoted to a text-based Agent architecture that guarantees reliable advice.

Accomplishments that we're proud of

Successfully building a "MetaEngine" that transforms raw API data into actionable game logic for an LLM.
Implementing a complex LangGraph workflow that seamlessly switches between internal database queries and external web searches.
Creating a tool that genuinely democratizes "Challenger-level" knowledge for casual players.

What we learned

Data Quality is King: We learned that even the smartest LLM cannot compensate for messy data. A significant portion of our effort went into cleaning and mapping the Riot API data to make it "AI-ready."
Agentic Reasoning: We gained deep hands-on experience in controlling LLM behaviors—teaching the model how to think like a gamer, rather than just retrieving text.

What's next for Teamfight Tactics's Tutor for newbie players.

The vision is to close the loop on our original challenge: Computer Vision. We plan to implement a custom OCR (Optical Character Recognition) layer to read the game state directly from the screen, removing the need for the user to type. Ultimately, I aim to evolve this project into a scalable "AI Co-Pilot Platform" for competitive gaming, turning complex data into simple, winning decisions for everyone.