I Built a Quotation Generator for Kenyan Street Welders Using Gemma 4's Vision

How I Used Gemma 4

I went with Gemma 4 26B A4B, the Mixture-of-Experts variant. Three reasons.

Vision that understands intent, not just pixels. The model doesn't just see "a rectangle with vertical lines." It recognizes that's a gate design with panels, estimates it at roughly 12 feet wide, and knows that means four hinges, not two. It reads hand-drawn sketches and blurry phone screenshots equally well.

Reasoning that produces consistent numbers. This isn't image captioning. The model estimates material quantities, looks up realistic Kenyan prices (square tubes at KES 3,500, electrodes at KES 1,500, fundi day rate at KES 1,200), then adds everything up correctly. The grand total actually matches the sum of its parts.

MoE efficiency for real-world speed. 26B total parameters, but only 3.8B active per token. A fundi standing in front of a customer doesn't want to wait 30 seconds. The MoE architecture gives near-31B quality while keeping response times practical on a mobile connection.

The technical setup is minimal:

Photo upload → FastAPI backend → Gemma 4 (Google AI Studio API)

→ Structured JSON response → Rendered quotation → PDF or WhatsApp

The system prompt tells Gemma 4 to act as a Kenyan construction estimator with current market prices. No RAG, no database of prices, no fine-tuning. The base model already knows enough about Kenyan hardware stores to produce quotations that a real fundi would look at and say "yeah, that's about right."