From the build
Dev Blog
A running log of what we're building and why.
The Signal, the Noise, and the 10-Day Data Fetch
PRATE was the wrong covariate. Fixing that cost 10 days of data fetching, and the results still weren't good enough — which led to a bigger pivot than I expected.
Read more →The 7.3% Problem: What's Hiding Inside the 79% Baseline
Fine-tuning Chronos-2 made my outputs worse — and that regression forced me to ask a harder question about my 79% baseline. The answer wasn't pretty.
Read more →I Thought My Ingestion Pipeline Survived a 4-Hour Outage. It Didn't.
My NOAA ingestion script hung silently through a mountain internet outage. The timeout fix was easy. The recovery working flawlessly was the real win.
Read more →The 79% Wall: Pivoting to Fine-Tuned Chronos-2 for the PowDay MVP
My initial inference engine hit a 79% accuracy ceiling on zero-shot Chronos-2. Here's what the failure mode revealed, and why I pivoted to fine-tuning on 10 years of Sierra Nevada data.
Read more →Why I Built PowDay.AI
Tahoe's micro-climates are genuinely hard to model. When I started learning about time-series AI, I saw a possible angle.
Read more →