Royal Coast Caddie drafts. The instructor decides.
Caddie is an AI workspace for college instructors. It drafts grading feedback against your own rubric and shows the passage behind every score. Student names and personal details are removed before any AI model sees the work, and you approve every word before a student reads it.
What it does
Drafts a score and feedback for each rubric criterion, with the passage from the student's own text that justifies it.
What it protects
Names and identifying detail are replaced before anything is sent. The instructor sees the original; the model never does.
What it costs you
A short interview to build the rubric once, then reading and approving drafts instead of writing them from nothing.
Caddie is in development and is not yet available. We are looking for faculty to help shape it.
The problem
Right now the choice is time or privacy.
Faculty are already using general-purpose AI to grade, often by pasting student work into a consumer chatbot. It saves real hours on the task that eats the most of them.
Institutions cannot approve that, because student work leaves their control the moment it happens, with names attached. So instructors choose between the help and their students' privacy, and nobody is comfortable with either answer.
Caddie removes the choice. The help arrives, and the identities never leave.
Built for higher education
Built for higher education.
Graduate seminars, case analyses, lab reports, and discipline-specific rubrics are the design center, not an afterthought to a school tool. Caddie learns the shape of your course rather than the subject. It does not need to know chemistry. It needs to know what a chemistry lab report looks like.
The privacy vault
Identity never reaches the model.
The protection is a property of how the system is built rather than a promise in a contract. Student work passes through a vault that replaces identifying detail before anything is sent, and the only route to an AI model refuses any text the vault has not signed.
What you see
Every Saturday of high school I woke at four to help Abuela Rosa open her bakery on Bleecker Street in Utica. That certainty ended the spring my brother Daniel was diagnosed with leukemia.
What the model sees
Every Saturday of high school I woke at four to help PER-KYMF open her bakery on a street in a mid-sized city. That certainty ended the spring my brother PER-86FE was diagnosed with a serious illness.
PER-KYMF a token, standing for one person a serious illness a detail generalised, not removed
Four layers, all running on your own machine
Detection runs entirely inside the boundary we control. Nothing in this step costs an API call, and nothing in it leaves that boundary.
Your roster
Exact matches first, then accents, nicknames and misspellings. The most reliable signal, and the one nothing else is allowed to override.
Formats
Emails, phone numbers, student IDs, and addresses. Validated rather than pattern matched, so a page range is not mistaken for an ID.
Names and places
The people a roster cannot know about: a landlord, a sibling, a hospital, a high school.
The indirect ones
A hometown, a parent's employer, a family illness. Details that identify a student without naming them, which is the hardest part and the part most tools ignore.
When Caddie is not sure, it asks you. A detail the vault cannot confidently place is held, and the essay is not graded until you decide. You see the sentence, the reason it was held, and a suggested replacement. You can accept it, cut the detail, or send it as written. Whichever you choose is recorded.
Your judgment, not ours
Caddie drafts. You decide.
Your rubric, in your words
A short interview, informed by your syllabus, turns what you actually look for into criteria. It only asks what your syllabus could not answer, and it asks the questions nothing else can: what separates excellent work from adequate here, and what should not count against a student.
Evidence for every judgment
Each score comes with the passage that justifies it, highlighted in the student's own text. A quotation that does not appear in the submission is flagged rather than trusted, so the evidence is checkable rather than decorative.
It says when it cannot tell
A criterion it cannot assess from the submission is marked unassessed, not guessed at. Where the vault's edits may have cost a judgment something, the draft says so and points you at the original.
Nothing reaches a student or a gradebook without your explicit approval. Every data movement and every approval is written to an append-only log that an institution can review.
Where it stands
Honest about the stage.
Caddie is a proof of concept. It is not publicly available and it is not in any classroom. Here is exactly what runs today and what does not, because a tool that asks instructors to trust it should start by being accurate about itself.
Working
The privacy vault, all four layers, running locally
Working
Tokenization, re-identification and the dual view
Working
Grading drafts with evidence, against your own rubric
Working
The rubric interview, built from your syllabus
Working
Held-detail review and the append-only audit log
Not yet
Study materials, lesson plans and slide decks
Not yet
Canvas and other learning management systems
Not yet
Institutional sign-on and admin controls
Detection quality is measured against fixed targets on a test corpus before any real student work is involved. Those targets were set before the first measurement rather than after, and we will publish what we find, including where it falls short.
Design partners
Help us find what is wrong with it.
We are looking for faculty who teach writing-heavy courses, grade real stacks of work, and would use an early version on a real assignment. You would be working with something unfinished, which is the point: what you find now changes what gets built.
Or write to hello@royalcoastlabs.com with "Caddie design partner" in the subject.
If you are reviewing this for an institution
Caddie is designed to support an institution's obligations under FERPA by keeping student identities inside a boundary the institution's vendor controls, rather than by contractual promise alone. That is design intent, stated before any legal review, and we will say so plainly until a review is done.
The architecture document, the threat model, and the containment tests are available on request, along with a walkthrough of what crosses the boundary and what does not. Ask for the technical brief.