By Liliana Lopez

Over the past several weeks, I've had the opportunity to participate in a valuable learning experience at Definity: Building in the Age of AI, an internal workshop series led by our Managing Partner, Omar Del Rio.

This course challenged us to rethink how software gets built. The premise was simple: What happens when AI becomes a true engineering collaborator instead of just a coding assistant?

To answer that question, each participant built a real application from scratch using a structured, AI-assisted development workflow. My project was MetroRuta, an offline-first mobile app that helps commuters navigate Monterrey's metro system—even without an internet connection.

The application itself was rewarding to build. But the biggest takeaway wasn't the finished product; it was learning a fundamentally different way of engineering software.

1.      Building Software with a Process, Not Just Prompts

One of the first things Omar emphasized during the workshop is that AI is most effective when it's given structure.

Rather than jumping straight into code, every feature moved through a disciplined workflow we call GSD (Get Sh*t Done):

• Discuss the problem

• Design the user experience

• Create a detailed implementation plan

• Execute with specialized AI agents

• Verify and test

Each phase produced concrete artifacts before moving to the next. Instead of treating AI like autocomplete, we treated it like a team of specialists—researchers, planners, reviewers, implementers, and testers—all working together under human direction.

Rather than spending hours debugging architectural decisions after code was written, we debated them before writing a single line. Across the first three phases of MetroRuta alone, we resolved 31 architectural decision points before implementation even began.

2.     The Challenge: Build an Offline Transit App

MetroRuta was intentionally ambitious. The goal was to create an application that could:

• Route passengers across Monterrey's metro network

• Work completely offline

• Calculate walking directions to and from stations

• Present the three best route options

• Run on both Android and iOS

Because commuters often lose connectivity underground, relying on cloud APIs wasn't an option. Instead, the application uses:

• Graphology to model the metro as a weighted graph

• Bidirectional Dijkstra for route calculation

• MapLibre with self-generated vector tiles for offline maps

• Valhalla for pedestrian routing

Every design decision centered around one constraint: No internet required.

3.     AI Didn't Write the Software. It Challenged My Thinking.

One misconception about AI-assisted development is that the model simply generates code faster. That wasn't my experience.

The AI was most valuable when it challenged assumptions I didn't realize I was making.

During development, it helped uncover subtle issues that would have been incredibly difficult to spot manually, including:

• Coordinate ordering inconsistencies between [lat, lng] and [lng, lat]

• Incorrect transfer logic caused by assuming the first line in a station definition instead of deriving the correct line dynamically

• Route comparison issues where "dominated" routes appeared competitive until Pareto filtering was introduced

These weren't syntax errors; they were correctness problems, and catching them early saved significant debugging time later.

4.    Human Judgment Still Matters

One of the biggest lessons from the workshop is that AI doesn't eliminate engineering judgment—it underscores its importance.

Every phase ended the same way: I installed the application onto a physical Android device, enabled airplane mode, used the application in realistic conditions, and documented what happened.

For the final completed phase alone, I conducted four rounds of user acceptance testing.

Those tests directly influenced product decisions that AI could never have inferred from code alone, including:

• Adding a dedicated route detail screen

• Improving station labeling on the map

• Refining the navigation experience based on real-world usage

The AI generated ideas. Device testing identified which ones actually improved the product.

5.     The Results

Today, MetroRuta has completed three of its four planned phases, with the core experience fully functional and validated on physical hardware. The remaining crowdsourcing phase is paused while reviewing Mexico's updated GPS consent requirements under the LFPDPPP 2025 regulations.

6.     Building in the Age of AI Is About Better Engineering

There's a lot of conversation today about whether AI will replace software engineers.

After participating in this workshop, I think that's the wrong question.

The better question is: How do great engineers build differently when AI becomes part of the team?

For me, the answer is:

• More planning before implementation.

• More explicit decision-making.

• More testing.

• More documentation.

• Faster iteration.

• Better software.

AI didn't replace the engineering discipline; rather, it amplified it. That's what makes Building in the Age of AI so valuable. It’s about adopting a new operating model for software development—one where human judgment, structured processes, and AI capabilities work together to build better products faster.