Back to Home
Case Study 2026

AI Development for Product Managers

Ditching the static slide decks for a living prototype: here is what happened when I tried to build a complex system using AI agents.

After nearly two decades in the UK, I finally made the permanent move back home to South Africa. My entire life is currently packed into a cargo ship somewhere off the coast of Liberia, so I have been killing time at my mom’s house. Naturally, I ended up hijacking her laptop to mess around with Google’s new IDE, Anti Gravity.

I had played with GenAI before. At KPMG, I once used Gemini to whip up a REST API spec to get around a tight resourcing bottleneck. To my amazement, it actually made it into production. But those were small, isolated wins. This time, I wanted to see if an AI agent could actually help me build a real, living system from the ground up.

The first step was picking a stack. The language I actually started my career programming in is pretty much ancient history now, so I narrowed it down to JavaScript or Python. They both have massive communities and endless Stack Overflow threads, which is just as helpful for an AI's brain as it is for mine.

The Concept

The Problem

Anyone who has worked in logistics knows the headache of warehouse intake. A massive shipment arrives and suddenly nothing matches. The tags, the barcodes, and the paperwork are all out of sync, which grinds everything to a painful halt.

The Solution

I kept thinking: what if you could just feed a messy product identifier, like a scuffed barcode or a vague description, into a system and it automatically hunted down everything the internet knows about that product?

I mapped out a self-serve platform where you create an API endpoint and define the exact fields you need, such as material or dimensions. Your input goes straight to Gemini with a tailored prompt. The AI scours the web, intelligently maps that data back to your custom fields, and spits out a clean JSON response.

The Build Experience

I did not want to build a simple "Hello World" app. I wanted to test this with some real-world weight. The final system included Clerk Authentication, a Postgres database on Supabase, Google Gemini integration, and full management for organizations and custom APIs.

The "Echo Technique"

Since Anti Gravity is so new, there are not many guides available for complex tasks. To get over the blank-page syndrome, I used what I call the "echo technique." I treated Gemini like a sounding board, bouncing ideas back and forth until we had a solid plan and a tech stack. I used that same process to fine-tune the core prompt later on.

Co-Piloting with the Agent

Instead of going full sci-fi with an autonomous network of agents, I kept things grounded. I loaded our plan, the database scripts, and a few code snippets straight into Anti Gravity. After a few minutes and a hefty chunk of tokens, the agent handed over an implementation plan.

We did not just blindly follow it. Leaning on my background in product and architecture, I guided the AI step-by-step, sequencing the build exactly how I would run it with a team of human engineers:

  • 01. We started with the plumbing: wiring up the Postgres DB on Supabase and making sure the connection strings worked.
  • 02. Next was security and the basic UI. Once Clerk Authentication was running, we could actually log in and see data sticking to the database.
  • 03. Then came the heavy lifting: the logic for custom APIs. The AI nailed this user story with only a few minor tweaks.
  • 04. Finally, we added a testing panel so users could check the accuracy of Gemini’s results in real time.

Real Talk on Velocity

By the time I came up for air, I had spent about two and a half days on the build. Looking back at the pace of teams I have managed, this would have easily taken four or five developers a solid week. Like it or not, AI is going to fundamentally reshape how we work.

What to watch out for

The AI's pure enthusiasm to solve your problems is actually a bit of a security flaw. It wants to help so badly that it will routinely copy your live connection strings and passwords right into public test files. If you are not paying close attention, those secrets go straight to GitHub. This happened to me at least five times, so keep your eyes peeled.

My Execution Playbook

Start with standard chat Use a regular LLM interface to talk through your ideas and build a solid plan before you ever give the agent context.
Be ruthlessly explicit Spell out exactly which auth providers and databases you are using. Include explicit schemas to keep the AI from improvising.
Question the blueprint Do not take the AI's first plan as gospel. Think about how you have sequenced projects in the past and reshape the plan accordingly.
Break down your stories Small, intentional increments are vital. It makes debugging much easier when things inevitably go sideways.

Final Thoughts: The PM Role

There is an old school of thought that says a PM who gets too technical is sacrificing their product strategy skills. Honestly, that was always a misconception, but today it is completely obsolete. In this new era, PMs who truly understand the engine can use AI to completely obliterate the gap between a concept and a product.

We are moving toward a world where a specification is no longer just a document, it is a functional prototype. For lean products, a single PM might just build the entire thing themselves. This shift challenges the idea that execution is secondary to strategy. To make sure your vision survives, you have to get your hands dirty in the build.

To thrive in this landscape, being an end-to-end product manager is no longer just an edge, it is the requirement.