I moved to Tahoe four years ago and immediately became the person people texted for ski conditions.
There's no shortage of weather forecasts for the Sierra. The problem is Tahoe's micro-climates are genuinely hard to model — Heavenly and Kirkwood can diverge by feet, not inches, on the same storm — and climate variability has made it a harder problem than it used to be.
When I started learning about time-series AI, I saw a possible angle. SNOTEL ground stations already measure hourly snowpack, temperature, and precipitation data across the Sierra. NOAA's HRRR model produces high-resolution atmospheric forecasts every hour. What if you fused those two signals through a covariate-aware foundation model and generated probabilistic forecasts — P10, P50, P90 — instead of a single number that pretends to know more than it does?
That question turned into PowDay.AI. I'm running Amazon Chronos-2 fine-tuned on 10 years of Sierra Nevada data on consumer hardware, producing 48-hour quantile forecasts for five Tahoe resorts — no cloud costs, no vendor lock-in.
Over the next few weeks I'll be sharing what I'm learning as I build it: model selection, data pipeline decisions, backtesting results, and where it still breaks. If that sounds interesting, follow along.