Protégé

Inspiration: Richard P Feynman

We believe that one can only truly understand something, if they are able to explain it so someone else.

The assumption is that there are many applications that help you study as if you are a student and the app is a tutor. We made an app where you are the tutor teaching an AI agent about notes that you upload, meanwhile there is a shadow agent that judges the corrections that you make and can then teach you.

What it does

Protégé turns learning around: instead of asking an AI to teach the user, the user teaches Watson. Upload study material, review Watson’s beginner-level understanding, and explain or correct it in chat. Watson updates relevant paragraphs, flags identified errors, and asks follow-up questions. Moriarty can compare Watson’s original and revised documents with the uploaded source and report errors the user may have missed.

How we built it

We built the app with Vite, browser-based JavaScript modules, and CSS. Uploaded material grounds Watson’s initial draft and the evaluation. Gemini generates Watson’s responses and paragraph updates, and Moriarty’s report. A Node.js test suite covers key evaluation, metrics, and document-update behavior.

Challenges we ran into

Gemini’s free-tier quota can interrupt generation, making a live demo less predictable.
Model-generated JSON and evaluations can vary, so we had to handle malformed or missing fields and avoid presenting defaults as measured results.
The evaluator needs to compare the source, Watson’s original draft, the user’s teaching, and Watson’s final document to assess missed errors.
We refined the interface to make chat the main correction workflow and remove confusing placeholder numbers and labels.

Accomplishments that we're proud of

This is a Minimum Viable Product
Built a working upload-to-teaching loop: Watson drafts, the user teaches through chat, and the document updates.
Added automatic error flagging and a Moriarty report that distinguishes missed errors, corrections, false alarms, and final accuracy when available.
Kept Watson conversational with follow-up questions while making manual red-pen selection unnecessary.

What we learned

The inversion is the product’s core idea; the AI itself is only the mechanism. We also learned to distinguish model estimates from measured metrics, ground feedback in the uploaded material, and be transparent when API limits or incomplete analysis affect results. Additionally, the need for reliable AI tokens for not only API calls but also for building of the app itself is integral to software development.

What's next for Protégé

Folders for specific subjects/courses that can be added
Login page/auth
Fix api key limits and remove the portal from mvp
Make font/themes more accessible s.t. users can change it to their preference
Create a loop that user can exit at any point so that the user can go through as many iterations that they want and this can serve as a source of blurting
Speed could be improved; avoid latency
Add different types of ‘teaching’ such as short-answer and multiple-choice quizzes that are to be marked by the user.
Currently the comparison and accuracy is calculated using AI but in the future, it could be implemented deterministically using sentence comparison.
Create a client-server and add a database system.
Have it hosted publicly.