an AI tutor that teaches through questions, then builds the learning artifact you need.
Studi keeps the lesson in one conversation. It asks the next useful question, then generates a runnable or interactive Spark when seeing, testing, or manipulating the idea would teach it better than another paragraph.
most AI study tools optimize for producing an answer. that is useful when you are stuck on a task, but it is a bad default when the goal is to understand the idea well enough to use it again.
I built Studi around a different loop: start from what the learner already believes, ask one smaller question, introduce a contrast, and only name the concept after they have noticed the pattern themselves.
Sparks are what make that loop more than prompt writing. when words are the wrong medium, the tutor can build a graph, scene, quiz, flash-card deck, or runnable code exercise for that exact moment without sending the learner somewhere else.
Studi asks one focused question at a time, adapts to what you say, and introduces the formal name only after you have found the idea yourself.
When text is not enough, the tutor can build an interactive scene, Desmos graph, quiz, or flash-card deck around the exact concept in the conversation.
Code and Test Sparks turn a concept into a small TypeScript, JavaScript, or Python workspace. You can predict, run, inspect the output, and repair visible failing checks without leaving the thread.
Each topic keeps its messages, attachments, and generated Sparks together, so you can return to a lesson without rebuilding the context from scratch.
Attach a problem set, lecture slide, or photo of your notes and Studi uses that material as the starting point instead of teaching from a generic prompt.
Accounts, plan-based usage, billing state, file storage, and failure recovery are part of the same experience, not a demo layer added around the tutor.
The tutor calls one generic create_spark tool with a Spark type and learning context. A typed skill catalog supplies the right prompt, schema, and validator for scenes, graphs, code, tests, quizzes, and flash cards, so adding a new Spark does not require rewiring the agent.
A separate worker generates strict structured output, then Studi normalizes and validates it against the Spark contract before anything is stored or shown. Generated code never becomes trusted UI just because a model returned it.
Scene Sparks run as compact HTML, CSS, and JavaScript inside a restricted iframe with a CSP and no network access. A small runtime bridge reports readiness, size, interactions, checkpoints, and errors back to Studi without giving the scene access to the app.
Code and Test Sparks are admitted server-side, checked against plan limits, and run in isolated provider-backed workspaces with bounded files, commands, languages, and timeouts. The learner sees useful output without exposing production infrastructure.
Convex Agent owns the thread, messages, tool calls, and tutor behavior. Replies stream into the same persisted conversation through live subscriptions, while Clerk identity and usage state stay enforced on the backend.
Tutor replies and Spark workers use task-specific model routes with guarded fallbacks. Providers can change without rewriting the chat UI, artifact contracts, or the teaching flow.
runs on
hover a tool to see why it is here
seen enough:
back home