Full Stack Developer

A football, an image recognition model, three APIs, and a surprisingly large amount of debugging.

During my time as a Full Stack Developer at Dom Ventas Global, I was given a challenge that combined two worlds I had previously explored only in isolation: AI and full stack development.

My task was to build the complete marketing website for our sister company, BroadwingLabs.

The website had two responsibilities.

First, it acted as the public face of the company, showcasing the marketing services we provided across marketplaces such as Amazon, Bol, Noon, and TikTok.

Second, it included an experimental AI tool that attempted to answer a simple question:

"If I upload an image of a product, which advertising keywords should I target?"

Simple questions have a remarkable ability to become complicated very quickly.

The Idea

Suppose a user uploads a photograph of a football.

Instead of manually searching for advertising terms, the system would automatically identify the product and generate marketing insights.

The workflow looked like this:

The system began by passing the uploaded image through a locally hosted vision model to identify the object.

Once the object was recognized, the generated description was used as a search query for the Amazon Search API.

The resulting products were converted into ASINs, which were then passed to the Amazon Ads API to retrieve relevant advertising keywords.

The final output was a collection of marketing terms that could help advertisers optimize their campaigns.

This Was Hand-Coded, Not Vibe-Coded

One detail that deserves a special mention is that this entire project was hand-coded.

Every API integration, database query, frontend component, backend endpoint, and AI workflow was implemented manually. There were no AI coding assistants generating components, suggesting architectural patterns, or automatically wiring together APIs.

Looking back, I'm actually grateful for that.

Building the system from scratch forced me to understand every stage of the pipeline. I couldn't treat the AI workflow as a black box because I was responsible for every moving piece, from the frontend and backend to the vision model and Amazon APIs.

Was the implementation perfect?

Absolutely not.

• There were no retries.
• There were no fallback mechanisms.
• There were no guardrails.

If an API failed, the workflow failed. If the vision model misidentified an object, the entire pipeline followed that mistake.

But that's exactly why this project remains one of my favorite learning experiences.

Every bug became a lesson. Every failed request became an opportunity to improve the system.

Today, AI-assisted development tools can generate significant portions of an application in minutes. This project belonged to a different era in my journey.

It was built the old-fashioned way: One function, one API call, and one debugging session at a time.

Beyond AI

The project wasn't limited to AI.

I was also responsible for developing the complete website from the ground up.

The frontend was built using Vue.js and Vuetify, while Flask and PostgreSQL powered the backend.

The website also functioned as a landing page for businesses interested in our marketing services across multiple e-commerce platforms.

What I Learned

This project taught me several important lessons.

AI models are only one component of a production system.

The real challenge lies in orchestrating models, APIs, databases, and user interfaces into a single, reliable workflow.

More importantly, I learned that a working prototype and a production-ready AI application are two very different things.

And sometimes, all it takes to discover that difference is a picture of a football.