When I took Science and Strategy of AI at UCLA Anderson, I knew I wanted my final project to be a gaming application. TFT was the obvious choice — it's the game that has kept me hooked across multiple sets since COVID, even though I've only been able to grind to Emerald.

After learning about retrieval-augmented generation and how models can be grounded in domain-specific context, I wanted to build an AI that could reason about TFT the way a high-elo player does, rather than just recycle generic tier-list advice. High-elo players do not win because they memorize patch notes. They build intuition through thousands of games: when to greed, roll, or level; how to preserve streaks; and when to slam items versus hold for BIS. That kind of knowledge usually lives in experience and VOD review, not in a system that can explain the gap back to a player in a structured way.

This project is my attempt to close that gap. It combines curated high-elo match data with LLM-based reasoning so the output feels less like a static stats page and more like a coach that can explain what winning TFT actually looks like.

The system is built around a dataset of ~235k Challenger and Grandmaster TFT matches from North America, last updated on June 19, 2026. From that data, I compute benchmarks around leveling efficiency, economy management, placement outcomes, and item performance — with win-rate metrics anchored to top-4 finishes so the model is grounded in what successful play looks like, not what's merely common.

The reasoning layer uses a RAG-style approach. Instead of relying only on what a model may have seen during training, each response is grounded at inference time in a TFT-specific knowledge base — champion data, traits, item stats, and high-elo benchmarks — which keeps the system tied to current game context and structured evidence rather than generalized recall. Three models work together: Gemini 2.5 Flash handles coaching and live search, Groq LLaMA 3.3 handles champion-specific tips and fallback responses, and Riot's Match API provides personal game history when available.

Coaching also adapts over time using a multi-armed bandit. Each session, a UCB1 algorithm selects one of five focus areas — economy, comp flexibility, item prioritization, board consistency, or meta alignment — and steers the AI's emphasis toward whichever has produced the most placement improvement for that specific player. The bandit balances exploration (trying all focus areas) with exploitation (leaning into what has worked), so the coaching gets more targeted the more you use it.

This tool was built as an academic project. Its data comes from third-party sources and has not been fully vetted, so it should not be treated as production software. It currently relies on a personal Riot API key, which means personal game lookup can break when the development key expires; those keys require manual renewal on a short cycle. A production Riot API key is the next step needed to support a reliable, production-ready version of the tool. Champion and item references also depend on CommunityDragon and Riot Data Dragon, so some numbers may lag the latest patch.

The Riot API provides an end-of-game snapshot for each match, which is enough to analyze board strength, itemization, placements, and outcomes at a meaningful level, but not enough to reconstruct every in-game decision. That makes it well suited for personal match analysis, while still falling short of what a true replay feed or event-level telemetry would allow.

Because the benchmarks come from Challenger and Grandmaster games, the gap between your decisions and the reference numbers can look large. That is expected: high-elo players operate at a different speed, with different habits and a different meta than most of the ladder. TFT Coach is designed to turn that gap into specific, actionable steps you can use to improve from your current rank.