The Problem Behind the Screen
A customer approaches a welder in Nairobi, holding up a phone with a Pinterest screenshot. “I want this gate. How much?” The welder squints at the image, performs mental calculations, and guesses a number. He scribbles it on a scrap of cardboard. No material breakdown, no line items. The customer leaves, unsure of the price, and finds someone who appears more professional.
This scene repeats thousands of times daily across Kenya’s jua kali sector. Fifteen million informal artisans lose work simply because they cannot produce a professional kenyan street welder quote on the spot. That is why I built a tool to bridge this gap.
How I Solved the Quote Challenge
Jua Kali Quote is an AI-powered quotation generator designed for Kenyan artisans. It takes a photo of any reference image — a sketch, a Pinterest screenshot, or a photograph of an existing gate — and produces a structured quotation with materials, quantities, and pricing in Kenyan Shillings. Here are the seven ways I built it to serve fundis working under the hot sun.
1. Choosing the Right AI Model for Vision and Reasoning
The core of the tool is Gemma 4 26B A4B, a Mixture-of-Experts (MoE) model accessed via Google AI Studio API. I selected this model for three critical reasons. First, it understands visual intent rather than just pixels. When the model sees a rectangle with vertical lines, it recognizes that as a gate design with panels, estimates its width at roughly 12 feet, and calculates that it needs four hinges. It handles hand-drawn sketches and blurry phone screenshots equally well.
Second, the model produces consistent numerical outputs. This is not simple image captioning. The AI estimates material quantities, looks up realistic Kenyan hardware prices — square tubes at KES 3,500, electrodes at KES 1,500, and the fundi day rate at KES 1,200 — then sums everything accurately. The grand total actually matches the sum of its parts.
Third, the MoE architecture delivers practical real-world speed. With 26 billion total parameters but only 3.8 billion active per token, the model returns responses fast enough for a fundi standing in front of a customer. A kenyan street welder quote appears in under a minute, even over a mobile data connection.
2. Building a Minimal Backend with FastAPI
The technical stack is deliberately lean. Python and FastAPI handle the backend. When a user uploads a photo, the backend sends it to the Gemma 4 API along with a system prompt that instructs the model to act as a Kenyan construction estimator with current market prices. The AI returns a structured JSON object containing materials, quantities, unit prices, labor estimates, and a grand total.
FastAPI processes the JSON response and sends it back to the frontend as a formatted quotation. The entire cycle — upload, API call, response, render — takes about 30 seconds on a typical mobile connection. No database of prices exists. No retrieval-augmented generation system sits behind it. The base model already knows enough about Kenyan hardware stores to produce realistic figures.
This minimal setup means the tool runs with a single API key and a basic server. Any fundi with a smartphone and an internet connection can access it through a web browser.
3. Designing a Mobile-First Frontend with Simple Technologies
I built the frontend with vanilla HTML, CSS, and JavaScript. No frameworks, no heavy libraries. The interface loads quickly even on older Android phones with limited processing power. The layout works in portrait mode, with large buttons and readable fonts sized for outdoor use under bright sunlight.
The homepage presents a simple workflow: upload a photo, select the trade (welding, carpentry, masonry, plumbing, electrical, or painting), and click generate. The quotation appears as a clean list of line items. Each line shows the material name, quantity, unit price, and subtotal in KES. The fundi can tap any line item to adjust a price they know better from their local supplier. The totals recalculate live as they edit.
This design respects the fact that fundis work on their feet, often with customers watching. They need speed not complexity.
4. Handling Unclear or Multi-Angle Reference Images
One of the biggest technical challenges involved handling poor quality reference images. Customers often show screenshots that contain multiple gates, partial views, or hand-drawn sketches with uneven lines. The Gemma 4 model processes these images intelligently.
When a Pinterest screenshot shows three different gate designs, the model isolates the most prominent one and generates a quotation for that design. If the image shows a gate from an angle, the AI estimates the dimensions using standard proportions. For hand-drawn sketches, it interprets rough lines as specific structural elements — vertical lines become uprights, horizontal ones become crossbars or rails.
The system prompt includes instructions to make reasonable assumptions and note them in the quotation. For example, if the image does not clearly show hinges or locks, the AI defaults to standard hardware and marks it as estimated. This transparency builds trust with both the fundi and the customer.
5. Enabling Editable Line Items with Live Recalculation
A fixed quotation rarely works in the jua kali sector. Material prices fluctuate weekly. A fundi might prefer stainless steel over mild steel for a coastal client. Labor rates vary by neighborhood — KES 1,200 per day in one area, KES 1,500 in another. The tool must adapt to these real-world variables.
Every line item in the quotation is editable. The fundi taps a material price, types a new figure, and the subtotal updates immediately. If they switch from mild steel square tubes to stainless steel, they adjust the unit price and the total recalculates. The same applies to labor: if the job requires two days instead of one, they double the labor line and the grand total changes in real time.
You may also enjoy reading: 7 Buzzworthy Films for World Bee Day (Oscar-Robbed Included).
This flexibility transforms the tool from a rigid estimator into a negotiation aid. The fundi can show the customer how different material choices affect the final price, building confidence through transparency. A clear, editable kenyan street welder quote helps both parties agree on the scope and cost before work begins.
6. Adding a Confidence Indicator for Trust Building
One feature I prioritized is the confidence indicator. Every quotation displays a range: “materials ±15%, labor ±20%.” This signals honestly that some numbers are estimates, not fixed prices. A firm but wrong number damages trust more than an honest range does.
The confidence indicator serves three purposes. First, it protects the fundi. If material prices rise between the quote and the purchase date, the 15% buffer covers the difference. The fundi does not eat the loss. Second, it reassures the customer. Seeing a range feels more transparent than seeing a flat number that later changes. Third, it creates a conversation point. The fundi can explain: “The steel price might shift by next week, but I will give you the actual receipt and adjust the final bill accordingly.”
Customers who have been burned by vague quotes in the past appreciate this honesty. They see a WhatsApp message with line items, materials, and a confidence range — and they decide to hire the fundi on the spot.
7. Using Local Storage for Quote History and Repeat Business
Every quotation is saved locally in the browser using localStorage. No cloud account required, no internet needed to view past quotes, no monthly subscription. The fundi builds a searchable history of every job they have quoted. They can reference a previous gate design when a new customer wants something similar, copy the base quote, adjust the dimensions, and generate a new quote in seconds.
This history acts as a portfolio. A fundi who has quoted twelve gates over six months can show a customer their past work, all neatly saved with dates and materials. Repeat customers receive consistent pricing because the fundi can pull up the original quote and apply the same rates. The local storage also helps with inventory planning — a fundi sees that they have quoted three gates this month and can estimate how much steel to order from the supplier.
The trade-off is that quotes only exist on the device where they were created. If the fundi changes phones, the history does not transfer. But for most fundis who use the same phone for years, local storage is practical and private. No one else sees their pricing data.
Beyond Welding: A Generalizable Pattern for Informal Trades
The same approach works for carpentry, masonry, plumbing, electrical, and painting jobs. A customer shows a photo of a cabinet from a magazine. The AI recognizes the joinery type, counts the shelves, estimates the wood volume, and generates a quotation with timber type, hinges, handles, and labor. A homeowner describes a wall they want to extend — the AI processes a sketch, calculates bricks, cement, and sand, and returns a masonry quote.
The real story here is a generalizable solution for informal trades that lack digital quoting tools. Six million micro-enterprises in Kenya could benefit from this pattern. The technology is accessible — an API key, a basic web server, and a smartphone. The barrier is not technical. It is the habit of scribbling estimates on scrap cardboard.
Imagine a welder who has worked in jua kali for ten years and is skeptical of AI. They try the app on a simple gate sketch. The breakdown matches their own mental estimate. They start using it for bigger jobs. A young fundi with no formal business training uses the app once, gets a positive response from a customer, and begins building a reputation for professionalism. A customer who has been burned by vague quotes before sees a WhatsApp message with line items and a confidence range and decides to hire the fundi on the spot.
These scenarios are already happening. The tool bridges perception and reality. It turns a fundi’s accurate mental math into something the customer can see, understand, and trust. That is the real value of a structured kenyan street welder quote.
