The Cost of Wrong Assumptions in AI Development

The Cost of Wrong Assumptions in AI Development

I’ve spent the last few years building AI-powered tools for everyday life in Kenya — from a logistics platform for sand and aggregate delivery to an e-commerce store that leverages AI, language learning apps focusing on Kenyan dialects, educational platforms tailored for children, and a planning assistant. Each product represents a unique challenge, but they all stem from the same root issue: assumptions we make early on can be costly when it comes to building with AI.

A Journey of Discovery

Building Arise & Shine Transporters was my first real foray into the world of AI-assisted logistics. The project aimed at solving manual inefficiencies in a sand and aggregate supply business, but I quickly hit a wall: language barriers were proving insurmountable given our current data constraints.

Initial Assumptions

I assumed that since Arise & Shine Transporters was a real-world problem with clear needs — GPS tracking for fleet management, cost-based pricing based on distance travelled — the challenge would be in finding or adapting existing AI solutions. I expected to find pre-built models or libraries that could handle these tasks efficiently.

The Unexpected Turn

But as it turned out, many of those “pre-built” solutions didn’t exist because they simply hadn't been applied to such specific African contexts where data availability was limited and varied wildly between regions. This meant the cost and complexity of developing custom AI models using conventional methods skyrocketed beyond what I had anticipated.

The Role of Claude and Codex

I reached out to Claude, a language model from Anthropic that can help craft prompts for various tasks including fine-tuning existing OpenAI models. Together with Codex (a tool by Anthropic), we were able to create customized AI solutions without starting from scratch. This not only cut down on the time and resources needed but also allowed us to incorporate specific linguistic nuances critical for real-world applications.

Insights from Arise & Shine Transporters

While Arise & Shine Transporters was a significant project, it’s not just one instance where assumptions led to challenges — they’re common in many of our AI-powered products. For example, with Local Dialect, I initially underestimated the complexity involved in creating an app for multiple Kenyan dialects that have different linguistic structures and usage patterns.

Finding Creative Solutions

To tackle these issues, we relied heavily on Claude’s ability to reason through language models and suggest effective strategies based on vast knowledge bases. We found that instead of trying to adapt pre-built solutions wholesale, it was more productive to use Claude and Codex as a bridge between existing AI tools and our specific requirements.

The Human Side

These challenges underscore the human element in building with AI: even when faced with seemingly insurmountable problems or limited resources, there’s always room for creativity and innovation. By embracing these solutions — whether they come from established models like Claude and Codex or by leveraging existing ones cleverly — we can overcome many of our initial fears.

Navigating Forward

As I look to the future with Arise & Shine Transporters and other AI-powered products, it’s clear that assumptions about what is possible will continue to shape our projects. But these insights from early challenges have given us a strong foundation for more efficient development processes in the years ahead. We’ve learned that sometimes stepping back to understand context better can lead not only to improved outcomes but also to smoother integration of AI into everyday solutions."

"AI model developer working with code editor