Assessor: privacy-first bulk AI marking
Load a rubric and a class set. Get back a marksheet you own, with the model never touching a name or a total.
Rubric in, de-identify, mark against the rubric, marksheet out. The teacher owns the totals; the model never totals.
AI product · assessment · data protection · live
The problem
Marking a class set against a rubric is slow, and student work is sensitive. Teachers want the speed without handing student data to a model, and without a black box that quietly invents a grade it cannot justify.
What I built
Assessor takes a rubric, a class set of work and a roster. It de-identifies every piece through the same isolated gateway that powers Kindred, marks each one against the rubric, declares plainly what it cannot judge, and hands back a live-formula .xlsx marksheet the teacher owns and can change. It is live on Vercel.
The part that matters
Code owns the totals; the model never totals. The AI judges against the rubric and flags anything it genuinely cannot assess as “not assessable”, rather than shipping a confident wrong mark. The de-identification spine is shared with Kindred at the config level, not by coupling code, so two products lean on one properly built privacy layer.
What it demonstrates
Productising a shared privacy core into a second product, clean architectural boundaries, and honest AI that is explicit about its limits, with a human owning the final call.
What I would bring to your project
If you want AI speed on sensitive, high-stakes work, I know how to keep the data safe, keep the maths in code, and make the model say “I am not sure” instead of guessing.