Field notes
Building Raychis in the open.
A 10-part set of field notes on building a real-world, on-device plant identifier from scratch, written while it was still uncertain whether it would work. What it produced isthe app's offline, on-device identification.
The complete series: all 10 parts published

Plant Identification Looks Solved. Until You Try to Build It
A plant app told me a toxic plant was basil. I spent a year building a plant identifier from scratch to understand why, and to test whether a different architecture could work.

The First Thing That Broke Was the Data
The model could not improve until the dataset could be trusted. That meant finding corrupt images, duplicate files, class imbalance, and the pipeline work underneath them.

The Data Looked Fixed. The Model Knew Better
The class distribution finally looked right, but the model still failed. The dataset was cleaner, not more informative, and training exposed the difference.

The Commit Before the Silence
The last commit of 2025, then fifty-two days of nothing. Burnout, stale metrics, and the first decision to stop building Raychis alone.

The Two Weeks I Stopped Building
Coming back to the project, getting the collectors running again, and spending two weeks writing the specification that would decide whether Raychis could keep going.

The Data Valley
The first real model reached 42.7% top-1 accuracy on 2,592 species. It was carefully logged, clearly unshippable, and the start of the valley before the final climb.

The Reframe
Reading papers in the valley, and the shift from training a classifier to training embeddings the app could actually search.

The Breakage
I went on holiday on April 3rd. While I was away, the ArcFace run reached zero validation accuracy and exposed the bugs hidden inside the reframe.

The Rule I Set for Myself: The Stage Gate
Accuracy ticking up, the clock ticking down, and the rule that kept me honest about what was actually ready to ship.

The App That Will Tell You It Might Be Wrong
Nine months after a plant app called annual mercury basil, an app that says when it is not sure. Where the series began, and where it lands.
The series is published first onMedium; these are the same pieces, kept here for the record.