Operating Philosophy: Technology
Technology Should Solve the Problem, Not Define It
Why AI architecture, governance, technical standards, and human approval should be designed around the problem being solved rather than the technology available.
Technology is the third pillar of my People, Business, Technology framework. It asks what the simplest reliable way to reach the outcome is, and where human judgment has to remain.
Technology makes it possible. It is the enabler, not the destination. That sounds obvious until you watch how many AI projects start, with the question "where can we use AI?"
That question has the order backwards. It starts from the tool and goes looking for a problem. I start from the other end: who is affected, what outcome matters to the institution, and only then, where does technology, including AI, create leverage? Sometimes the answer is a language model. Sometimes it is a clearer policy, a better-designed assignment, or a standard that already exists.
Standards before custom builds
When a standard covers the need, I favor it over building something new. Identity verification for secure assessment is a good example. Legacy facial-recognition proctoring carries real privacy risk, and it can flag a student as a cheater on a camera's say-so.
Established authentication standards, such as passkeys built on WebAuthn, can give strong sign-in assurance without a custom biometric pipeline. That means a smaller trust boundary, no biometric data leaving the device, and no custom model to maintain.
They have limits, though. A passkey confirms that the right credential was used. It does not, on its own, establish that the enrolled student is the person taking the exam from start to finish. So I treat standards like these as one layer of identity assurance, used where they fit, not as a replacement for it. The best technology decision is often the one that removes a system rather than adds one, as long as you are honest about what the remaining system does not cover.
Structure over promises
In a multi-tenant education platform, one misconfigured query can expose one institution's student data to another. That is the difference between a platform an institution trusts and one it cannot legally use.
So the standard I direct engineering toward is isolation enforced at the data layer, not only in application code that has to catch every case.
Compliance frameworks matter, but they are not the mechanism. FERPA sets legal obligations for how student records are handled, and a SOC 2 report attests to how a provider's controls are designed and operate. Neither makes a system secure by itself. Security comes from defense in depth: isolation, access controls, testing, and clear accountability, so protection does not depend on everyone being careful every day.
Visibility over surveillance
The same principle shapes how I think about academic integrity. Detection asks technology to deliver a verdict it cannot reliably support. Process visibility asks it to do what it does well: record how work developed, so a faculty member can read the context and decide.
One design hands judgment to a probability score. The other hands better evidence to the person who should be judging.
I build by the same rule
I hold my own tools to this standard. I built a product operations workspace where AI agents compile standups, analyses, and drafts from the systems I work in every day. The AI drafts. I review, correct, and decide. I built the workspace itself using Claude in VS Code under the same rule: the AI writes the code, and I decide what ships.
The most useful lesson from building it: when an AI mistake can be prevented through system design, I would rather enforce that rule in code than rely on a prompt asking the model to behave. Prompts help guide responses, but permissions, approval requirements, data access, and other critical boundaries should be enforced by the application wherever possible.
In my workspace, nothing reaches the issue tracker or a customer record without my approval, and each agent gets only the tools its role needs. Not every rule has made that move yet. Not every model error can be prevented either, which is why evaluations, human review, and clear recovery paths matter just as much.
Questions to ask before you add AI
- What problem does this solve, and for whom? If the answer starts with the technology, start over.
- Is there a standard that already does this? Building custom should be the exception you justify.
- What happens when it's wrong? Every AI system will be wrong sometimes. Design the path from error to correction before launch.
- Where does the human decide? Name the person and the moment.
Technology goes last in my order for a reason. Put first, it becomes the goal, and a goal of "use AI" is met the moment a model is switched on, whether or not anyone is better off. Put last, it answers to people and to outcomes.
In the end, any technical implementation should be judged against the people it serves and the outcomes it was designed to achieve.
Part of my People, Business, Technology series. Start with: Why AI Adoption in Higher Education Starts With People. Also read: How to Measure the Real Value of AI in Education.