BenchWarmer

Inspiration

The world of fantasy sports is built on a lie. We’re told that skill wins championships, but due to the dumb tools of today, wins and losses have simply come down to luck.

Current platforms like ESPN show you a single, pre-computed projection (like 14.5 points) with no context, no reasoning, and no personalization. This "black box" number is the same in every situation. Current dashboards don’t account for your opponent’s roster, upcoming schedules, or the fact that you’re 2–5 and need to take a high-risk, high-reward shot just to stay alive.

That forces every fantasy manager striving for the win to become a full-time data analyst — manually checking 10 different news articles and stat sources. If you’re not the one to see the 3 AM injury alert, scan the waiver wire, and make a blind decision, you’re bound to be left behind.

Thus, we were inspired to build the tool we’ve always wanted in fantasy football. We didn’t just build another dashboard, but a true autonomous AI agent that simplifies fantasy sports and gives you the upper hand.

What It Does

Simply put, BenchWarmer is your 24/7 AI co-manager.
It is a personalized agent that connects to your existing leagues and autonomously optimizes your roster in real time.

Autonomous Management

BenchWarmer doesn’t just advise. It executes lineup changes, scans the waiver wire, and places claims automatically, seconds after live events like injuries or benchings occur.

Proprietary Projections via Multi-Source Fusion

Our LightGBM model produces its own weekly projections using seven+ real-time data sources:

Sleeper API: live and historical stats
ESPN API: baseline projections and player context
The Odds API: real-time sentiment and implied scoring
Weather API: game-day environment
Historical Performance: prior season data trends
Opponent Data: defense-vs-position analytics

Explainable AI

Every choice is transparent. Our hybrid ML + RAG (Retrieval-Augmented Generation) pipeline explains lineup and waiver decisions in plain English.

Example:
“Starting RB Y: +1.4x matchup advantage vs. bottom-5 rush defense; Vegas O/U 51.5 → high scoring; Confidence 83%.”

Context-Aware Strategy

BenchWarmer adapts to your league’s situation:

Already clinched playoffs? It plans ahead.
Fighting for survival? It switches to boom-or-bust tactics to maximize your odds.

How We Built It

BenchWarmer runs on a five-service micro-architecture designed for scalability and ease of use:

MongoDB Atlas: our “source of truth.”
Stores all player stats, model outputs, and vector embeddings for memory and retrieval.
Data Aggregator (Python): continuously fetches and fuses live data from Sleeper, ESPN, and The Odds API.
ML Projections Service (Python): trains and serves a proprietary Random Forest model that powers weekly projections, retraining weekly on updated data.
AI RAG Service (Python): uses MongoDB Vector Search with LangChain to retrieve relevant player data and ground LLM outputs, producing factual insights for lineup and trade decisions.
Backend & Frontend (Node.js + Next.js) — user-facing chat and visualization layer for interacting with the AI co-manager.

Challenges We Ran Into

Our biggest technical challenge was accessing reliable historical data without enterprise-level API costs. Standardized APIs like Sportradar and SportsData.io restrict public-tier access, making them infeasible.

We found the perfect balance through the Sleeper API, and learned how to:

Loop through the NFL player list endpoint to get each player’s unique ID.
Use that ID to fetch week-by-week stats for every season.
Reference the state endpoint’s current week to keep data always up-to-date.
Merge these into one automated historical data pipeline for model training.

This turned a fragmented dataset into a dynamic, real-time database powering our ML engine — the key breakthrough that made full autonomy possible.

Accomplishments We’re Proud Of

End-to-End Proprietary ML Pipeline — our LightGBM model learns directly from our data, not ESPN’s recycled projections.
True Explainable AI — MongoDB Vector Search + LangChain RAG ensures every decision is grounded in factual data.
Real-Time Data Fusion — unified intelligence graph combining three live APIs (Sleeper, ESPN, Odds API).

What We Learned

In the AI era, your database is your agent’s brain.
Intelligence isn’t just about models — it’s about how data flows, stores, and communicates.

MongoDB Atlas served three essential roles: