Notes

Short writing from the edge between building and threat modeling

These are not archives for the sake of archives. They are the smallest useful pieces of thinking I want to keep public while the bigger work is still taking shape.

AI Security

K-12 AI rollouts fail at the boundary, not the demo

April 7, 2026

The flashy model behavior is rarely the real deployment problem. The real risk lives where student data, vendors, identity systems, and school operations touch.

Boundary failures are where the real risk shows up

Most school AI conversations start at the model surface: what it can do, how accurate it is, whether the prompt feels safe enough. That is useful, but incomplete.

The harder problem is boundary design. Who can upload what? Where does data move after the interaction? Which vendor gets retention rights? What gets logged, reviewed, or escalated when something goes wrong?

If those seams are weak, the product can look impressive in a demo and still be unsafe in a district rollout.

What I keep looking for

  • Places where product teams treat policy as a compliance checkbox instead of a design input.
  • Gaps between identity systems, vendor contracts, and classroom reality.
  • Interfaces that make risky workflows feel harmless because the friction is hidden.

That is why threat modeling matters early. It gives teams a map before the system starts lying to them with momentum.

Operating Notes

Use threat models before you promise the roadmap

March 24, 2026

A threat model is not just for security review. It is one of the cleanest ways to see which product promises are expensive, fragile, or irresponsible.

Threat modeling is a planning tool

Teams often wait to think about abuse, misuse, or system failure until there is already a roadmap, a prototype, and a narrative around launch timing.

That sequence is backwards.

If you model the trust boundaries first, you learn which features create operational drag, which ones need stronger permissions, and which ones are easy to explain but hard to govern.

The useful mindset shift

  • A roadmap is not just a list of features. It is a list of future obligations.
  • Every new system surface creates monitoring, review, and support work.
  • Some features are only good ideas if the organization can actually carry the control load.

That is a product question as much as a security question.

Prototyping

Video analysis prototypes need visible uncertainty

March 9, 2026

Vision systems become more trustworthy when the interface shows doubt, review states, and ambiguity instead of pretending every output is equally solid.

Confidence is part of the product surface

When people build video-analysis tools, the instinct is usually to maximize speed and make the output feel crisp. But crisp can be misleading.

If the system is guessing, the interface should help people understand that immediately. Not because uncertainty is elegant, but because it changes what a responsible user should do next.

Three things I keep coming back to

  • Confidence without context invites over-trust.
  • Review states are often more useful than raw model scores.
  • The best AI interfaces make room for correction without making the product feel broken.