Mazorca

Mazorca is an offline agronomic assistant built on Gemma 4 by a cocoa-growing developer in Cárdenas, Tabasco, Mexico, used for pest and disease identification and one-week management planning in cocoa plantations.

Mazorca uses a two-step diagnostic method: the first step observes with a neutral prompt that does not mention cocoa, disease or agriculture; the second step receives the first step's result and generates a diagnosis, and must not contradict it.

After the model output, Mazorca performs hard rule validation on the server: if no pod is visible, all pod-related findings and counts are removed; if the count far exceeds the observed estimate, it is scaled down proportionally.

Mazorca uses a zero-dependency Node implementation, using only node:* modules, requires no npm install, and can be copied to an old laptop via USB drive to run.

Before photo upload, Mazorca re-encodes via canvas and removes EXIF/GPS information, and all model text is inserted via textContent to avoid injection risks.

Mazorca's paying parties are cooperatives and purchasing centers, chocolate makers with sustainability or traceability programs, and extension technicians and agronomic consultants; individual farmers use it for free.

Developer hands-on testing found: Mazorca performs well on scene descriptions, but counting is an approximate estimate rather than an exact inventory; it once misread a pile of opened cocoa pods as 'overripe pods on the ground'.

Frosty pod rot has not yet been validated on Mazorca, because testing took place during the post-harvest period, when there were no photos of active frosty pod rot.

Mazorca runs as follows: ollama pull gemma4:e4b-it-qat, git clone the repository then npm start, and access it at localhost:3000; the model can be swapped via the MAZORCA_MODEL environment variable, the Ollama default timeout is 180 seconds, and the Kestra flow is an optional component that is disabled by default.

Mazorca's weekly action plan can be read aloud in Spanish via ElevenLabs.