Technology & research
How does a document become a question? When does a conversation need deeper work? When should a teacher take over? These notes explain our engineering choices and the research informing them.
Attendance recognition: three models, three distinct jobs
Vispo Care
YuNet, MiniFASNetV2 and SFace separate face detection, liveness evidence and identity matching.
From uploaded pages to editable questions
Vispo · OCR
Preserve page coordinates, group questions and align answers so every item can be checked against its source.
Classify how a student needs to respond
Vispo · Classification
Selection, text entry and drawing call for different interactions. Classification connects the document to the right response interface.
Fast / Slow Path: response, deeper work and handoff
Vispo · Systems
Not every turn needs the same depth of processing. Interaction signals guide escalation and define when to hand over.
When should a teacher step in?
Research · CMU
CMU's Lumilo research offers a way to think about persistent difficulty and the evidence teachers need before intervening.
How an educational chatbot supports thinking
Research · Tutoring
A correct answer is a starting point. Useful hints and student reasoning need deliberate design and evaluation too.
Separating questions and answers without losing their connection
Vispo · OCR
Numbering, sections and source positions connect each answer to an inspectable question and its original evidence.
AI question generation: from objectives to item-level checks
Vispo Research
Our question-bank research prototype plans, generates, checks and revises. Papers inform the method; teacher review determines whether an item belongs in a lesson.
Knowledge graphs: concepts, prerequisites and textbook evidence
Vispo Research
Make concepts and relationships inspectable, and give questions, materials and learning records a traceable context.
How we use research
We distinguish implementation, design direction and external research. Results from upstream models or papers are not Vispo outcomes, and a citation does not imply endorsement.
OUR VISION
Every interaction, a better understanding of the learner.
We want AI to learn from teacher–student interactions: to distinguish productive effort from persistent difficulty. By connecting questions, responses to hints and teacher feedback, we aim to better anticipate confusion and help teachers respond when support matters most.
Evidence for the decision. Agency for the teacher.
Training our own models and improving them iteratively is a research direction: turn teacher feedback and interaction signals into evaluation and training evidence, test how well models understand learners, then refine hints and the timing of alerts. Teachers retain control of instruction and intervention.