Howl Vision

Archived project

A second pair of eyes for everyone who loves an animal

Multimodal veterinary AI built on Gemma 4 for the Gemma 4 Good Hackathon (April–May 2026): offline triage in the browser, sub-second on-device image classification, and a full server pipeline pairing specialised vision models with Gemma 4 E4B.

The interactive demo that ran at this address has been retired now that the competition is over. Everything it was built from stays public and reproducible.

What is still available

• Source code on GitHub Full history from the first commit — PWA, FastAPI backend, vision service, benchmarks.
• Canine dermatology model EfficientNetV2-S, 6 classes, 94.0% accuracy on a held-out test set. CC-BY 4.0.
• Feline dermatology model EfficientNetV2-S, 4 classes, 90.1% accuracy. CC-BY 4.0.
• Blood parasite model Leishmania, Plasmodium and Trypanosoma in canine smears — 99.87% accuracy, n=3,032. CC-BY 4.0.
• Technical notebook on Kaggle Reproduces every claim below live on a T4 GPU, in about fifteen minutes.

What it measured

• 94.0%Canine dermatology, 6 classes, n=433
• 99.87%Blood parasites, n=3,032
• 675 msOn-device ONNX INT8, mid-range Android
• 19.7 MBQuantised classifier shipped to the browser
• 240 t/sGemma 4 E2B Q4_K_M via llama.cpp, 1.9 GB VRAM
• 0.0Fine-tuning delta vs base model — published rather than hidden

A note on the honest result

Three supervised fine-tuning runs on canine dermatology narratives produced a macro F1 of 0.873 — exactly the same as the base Gemma 4 E4B. A delta of zero. Rather than bury it, the project shipped the finding: for this task the base model was already sufficient, and the adapter added weight without measurable benefit. The decision not to fine-tune was the result of the experiment, not its failure.