AI is not a tool only for data teams. It changes how we design systems. It changes where architects sit. It changes what we measure.
Where Architects Sit in the Era of AI — A Clear Map for Software Leaders
AI is not a tool only for data teams. It changes how we design systems. It changes where architects sit. It changes what we measure.
I am Ermenildo Valdez Castro, Jr. Call me Ernie Castro. I have built systems at Google, Microsoft, Amazon, and Zulily. I have led teams that fought latency, tamed data, and pushed features at scale. I write plain. I write true. 🙂
Why architects matter now
Models are code. Models are product. Models are legal risks. A floating model breaks a system faster than a broken endpoint. Software architects who ignore AI will inherit chaos.
Keywords: software architect, AI architecture, machine learning, MLOps, model governance.
Where architects should sit
1) Near the data platform
AI eats data. If you are an architect, you must own data contracts. You must map data lineage. You must design pipelines that preserve truth.
At Amazon, we made a rule: if data could lie, we logged the why. That rule saved us from a bad recommendation run. I remember debugging a downstream service at 2 a.m. The trace pointed to a missed null check in a feature store. Small thing. Big failure.
2) At the model lifecycle table
Architects must design for training, validation, deployment, and rollback. Not just inference.
At Google, I sat in rooms where models were deployed like missiles. We built automated rollback paths. We tested for drift. The architecture was the safety net.
3) In platform and infra decisions
Choice of cloud, data lake, and orchestration shapes AI outcomes. Architects balance cost, latency, and reproducibility.
At Zulily, we learned to trade off speed for reliability. We postponed a flashy model to protect the checkout flow. The business thanked us the next quarter.
4) On governance and ethics
Design for transparency. Design for audit. Document assumptions. Keep humans in the loop.
At Microsoft, we ran ethical reviews for a search feature. The review forced us to add an explanation layer. The feature performed slower but stayed in production. That honesty built trust.
Practical lanes architects must run
- MLOps and CI/CD for models. Build pipelines that run with tests.
- Observability for models. Track accuracy and drift, not just latency.
- Data contracts and schemas. Guardrails prevent silent failures.
- Security and privacy. Encryption, access controls, and consent.
- Cost modeling. Training can spike bills. Plan for it.
A short checklist for leaders
- Own the contract between data and model.
- Treat models like first-class services.
- Build rollback and canary strategies.
- Instrument model performance and bias.
- Include compliance in design reviews.
A personal note
Once, at 3 a.m., I paged a team about a broken model. The fix was simple: a missing timezone conversion. I fixed it, pushed a hot patch, and wrote the postmortem. The team learned to assert time. That night taught me a lesson: architecture is humility. Know your assumptions. Test them.
Where architects add value
We add structure. We reduce surprise. We make failure survivable.
The Bible says build on rock, not sand. The advice fits architecture. Build on clear contracts. Build on repeatable pipelines.
Final words
AI will rearrange teams. It will demand new lenses. Architects will sit at new tables — the ones that own data, models, and trust.
If you lead, start now. Map the data. Define the contracts. Automate the tests. Protect the users.
I am Ermenildo Castro. Sometimes Ernie Castro. I have seen systems fail. I have helped them stand. Build for truth. Build to last. ✍️