# Whiff Scout

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Yu-Hao Yeh](https://cerebralvalley.ai/u/kekeken)
- **GitHub:** https://github.com/kenyeh727/mlb_data_research
- **Demo video:** https://youtu.be/884CkSqUSfs
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/353

Whiff Scout is a scouting-report agent: type a pitcher's name, get a data-backed scouting report. Three roles work in sequence. Scout plans the report. Data pulls 2026 Statcast. Analyst runs our own whiff model. Muse Spark then writes the prose from the model output, not from chat.

The demo video (1:56) shows the full loop on Jacob Misiorowski: 33.6% actual whiff vs 30.1% model-predicted, from 1,538 swings. Each report prints per-pitch whiff rates, left/right matchup splits, and the data window and game-type mix it covers, so a scout can see exactly what the numbers are based on.

For judges:
- Impact: the output is a paragraph a scout can use tonight, not a generic chatbot answer.
- Creative use of Meta tools: Muse Spark turns structured model output into scout prose, with limitations printed on the report.
- Demo: a 1:56 video, a public repo, and a ten-pitcher evaluation below.
- Technical depth: a gradient-boosting whiff model trained on strict swinging-strike labels, evaluated out of sample and on ten named pitchers.

Evaluation, actual vs model-predicted whiff, with swing counts:
- Jacob Misiorowski: 33.6% vs 30.1% (1,538 swings)
- Chris Sale: 28.1% vs 25.8% (1,505)
- Cristopher Sánchez: 28.8% vs 23.6% (1,657)
- Cam Schlittler: 26.9% vs 23.7% (1,747)
- Dylan Cease: 30.4% vs 27.6% (1,396)
- Logan Henderson: 23.2% vs 22.0% (901)
- Nolan McLean: 21.6% vs 19.4% (1,285)
- Parker Messick: 23.2% vs 22.0% (1,608)
- Payton Tolle: 25.0% vs 22.2% (1,354)
- Walbert Ureña: 26.7% vs 22.1% (1,253)

Mean absolute gap: 2.9 percentage points. The model underpredicts on all ten. That is a known, systematic miss, not a cherry-picked result.

Model: histogram gradient boosting, 75 trees, trained on 80,851 pitches (38,429 swings) with strict swinging-strike labels and a temporal split. Held-out AUC 0.777 (95% CI 0.766–0.789).

Limitations, stated plainly:
- The demo example in our earlier Rasmussen report mixes game types: 1,481 swings including 191 spring-training and 101 postseason pitches, so it is not regular-season-only. Overall 22.8% actual vs 21.9% model. Small pitch-type samples are flagged on the report: sweeper 15.4% actual vs 22.1% model on only 39 swings, curveball 24.0% vs 27.0% on only 25 swings.
- Coverage: 11 pitchers are cached locally. Other pitchers are fetched live from 2026 Statcast and then cached, which can take tens of seconds to a few minutes and needs the name spelled correctly.
- What's next: calibrate the model to fix the systematic underprediction, and add per-pitch swing counts to every report table.

Repo: https://github.com/kenyeh727/mlb_data_research (whiff-scout folder). Video link unchanged.

## More from Meta Global AI Developer Hackathon

- [MxA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/350)
- [CortexNET](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/351)
- [WalletShield](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/352)
- [MuseCell](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/354)
- [LibertyManzi(solo )](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/355)
- [Kareem](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/356)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/353. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt