This is an early look at FeedMeSports, where I’m testing AI + retrieval-augmented generation (RAG) to produce more useful weekly fantasy football picks. It’s focused on function, not design. A minimal demo is live at feedmesports.com.
Objective
Combine baseline betting data with fresh news to generate weekly picks that include why a pick matters: injuries, roster moves, and key storylines that could move outcomes.
How It Works (High Level)
- Baseline: Load market lines (over/under, moneyline) as the starting prior.
- RAG layer: Ingest daily news (injuries, depth chart notes, beat-writer reports), embed it, and retrieve only the most relevant snippets for each matchup.
- Adjust + summarize: The model explains the pick and confidence using the retrieved context (not generic chat), so users see the reasoning.
What the User Sees
- Weekly picks for each matchup.
- Short context: notable injuries, unit matchups, travel/rest quirks.
- A simple confidence indicator (low/med/high).
Why RAG?
Sports news changes by the hour. RAG keeps the model grounded in current information without retraining, and reduces hallucinations by forcing citations from retrieved articles.
Notes & Limitations (Current Build)
- Demo site is barebones (template UI) — focus is on the data flow and reasoning.
- Confidence isn’t a true probability; it reflects strength of evidence in retrieved sources.
- APIs can rate-limit during peak windows; fallbacks are still being tuned.
Next Steps
- Deeper injury/usage tracking (snap counts, target share, OL/DL health).
- Per-team embeddings so narratives persist week-to-week.
- Side-by-side view: baseline vs. RAG-adjusted, with links to the supporting articles.
Try It
You can see the weekly picks pipeline in action here: feedmesports.com (functionality > visuals; it’s a working test bed).