peakweek

peakweek answers one weekend question for the US Northeast and Mid-Atlantic: which trees near you are turning colour this weekend, and which day should you go out? Type a town, and you get one card for the coming weekend: for each of 8 common fall-colour trees, the chance it will be green, in colour or bare, a verdict per species (go, starting, wait or late), the best day in the next two weeks, and one line on how to recognise each tree. Print card gives a one-page black-and-white card to take with you.

The forecast is built from 6,650 iNaturalist community-science observations (September to November, 2018-2025) plus this season's weather, using TabPFN v2 open weights on a plain CPU with no GPU. The live site is rebuilt every morning by a GitHub Actions job that runs TabPFN once per 1-degree grid cell (103 cells) and publishes the cards to GitHub Pages; every card says which cell it is for and the day it was made. Settings were frozen before scoring: on the held-out 2025 year, TabPFN scored log loss 0.498 against 0.519 for logistic regression and 0.581 for climatology.

Vibe-coded evidence, in the author's own words on DEV (6 October 2026): "I built this with Claude Code, working as a small team of agents: one built the dataset, the features and the evaluation, and ran the TabPFN backtests on a small home server; another built the card, the page and the CLI." The post's AI disclosure adds: "this project was built with AI coding agents (Claude Code), which wrote the code, ran the evaluation and drafted this post." GitHub stars: 0 at the 28 October 2026 check (github.com/Bowen1314/peakweek).

Honest limits: forecasts are per 1-degree grid cell (about 110 x 80 km), so a cold hollow and a sunny ridge in the same cell get the same weather; the card only says a species is rarely recorded within 50 km when local iNaturalist counts are thin, rather than hiding it. Bare trees are under-photographed (only 347 of 6,650 observations), so the app is better at "not yet" than at "too late", and outside the covered region the page says so instead of extrapolating.