Breakneck Pace: Genesis with AI
We have figured out how to utilize the AI tools and improve development velocity on the enterprise products initially built during the Mesozoic era. It's a whole different ballgame when there isn't a mountain of tech debt to dig through.
 |
New Product
Luck has shined upon us and we got a new goal to build a new product with a clean slate. The last time we did this there were no AI tools available and the runway was measured in months (plural). Now we had one week.
AI-based rapid prototyping is a no-brainer. Product team AI-generated the requirements in a single day after a few meetings. Engineering managed to vibe code microservices and a web front-end by using widely available agentic models and came up with fully functional UI with back-end integration within a few days.
But here's the catch: the architecture isn't something to envy. Coding standards are non-existent. Testing has been done in a lab environment with mock data only. The real test will be getting this thing out to real customers and getting the feedback.
Looking down from the highest tree branch it all looks ideal: we managed to build something within a week's time and it "looks" production ready. No one dares calling it a "prototype". We may even see some revenue from this if we found the product market fit. Otherwise we have managed to fail-fast which is also a desirable outcome.
If we decide to make this a sustained product worth further investment it will be necessary to revise the architecture bottom up, to think about scalability, and to definitely rework some of the prototype code.
This is a pivotal moment and the future of the product hangs in the balance of the engineering team explaining the reality of AI-based rapid prototyping to the leadership team.
A stack of hands between all stakeholders means that everyone understands that a successful product prototype is a loan that has to be repaid if we are looking for a sustainable investment. The alternative is a continuation of rapid prototyping development and will result in creating a landfill full of future technical debt.
New Feature
From time to time we also got to have some fun with the existing enterprise grade products we owned that were actively used in production by high profile customers. We both implemented AI-supported features and used AI tools to rapidly develop a new feature.
The difference between a completely new product and an existing one that people are relying on for years is that all hell breaks loose if the latter misbehaves, or even worse, stops working altogether. No pressure. 😉
That doesn't mean we should be scared of using AI, but we needed to define some ground rules:
- No vibe coding
- Human-defined architecture; AI agent can help create the integration plan
- AI as an assistant only during integration; agents are acceptable for new isolated modules
- Four review stages before approving each PR: self, AI (a separate model), a peer, an architect
- Full unit testing coverage; rely heavily on AI here but also review
- Automated integration tests are an integral part of the definition of done
