Real-time voice and whiteboard lessons
LiveKit and OpenAI support streaming speech-to-text and text-to-speech, barge-in, and reconnect recovery. Interactive whiteboard lessons connect spoken tutoring with visual explanations.
Founder & Product LeadA voice-first assistant for student planning, coursework, and interactive whiteboard lessons, with source-linked memory and reviewable calendar actions.
As Founder & Product Lead, I am building NavoX, a voice-first AI assistant for college students. It connects weekly planning, calendar scheduling, coursework, and interactive whiteboard lessons through a persistent conversation, with encrypted source-linked memory and explicit approval before consequential actions.
LiveKit and OpenAI support streaming speech-to-text and text-to-speech, barge-in, and reconnect recovery. Interactive whiteboard lessons connect spoken tutoring with visual explanations.
Encrypted, source-linked memory preserves useful context while giving students retention, off-record, and deletion controls.
Google Calendar, Microsoft Graph through OAuth 2.0, and iCloud CalDAV adapters feed constraint-aware planning and scheduling.
Students review the exact proposed action before a calendar write. Supported Undo checks the resource version so a later edit cannot be silently overwritten.
An async FastAPI and PostgreSQL backend uses Temporal workflows for syncs, reminders, and briefings.
Student commitments are scattered across calendars, classes, and conversations. An assistant needs to recover from interrupted voice sessions, retain only the context a student chooses to keep, and distinguish a suggested schedule from an approved calendar write.
Authorize calendar providers and retain source-linked context under the student’s memory and privacy controls.
LiveKit manages real-time sessions while OpenAI powers speech and model interactions. Barge-in and reconnect handling make interruptions part of the interaction design.
Temporal coordinates syncs, reminders, and briefings. Approval state remains separate from background execution, and a materially changed action requires fresh approval.
PostgreSQL stores application state and source references with ownership boundaries. Encrypted memory has explicit retention, off-record, and deletion controls.
The recorded API/web regression checkpoint contains 5,985 tests, alongside strict mypy and Ruff checks. LLM evaluations cover schema validity and prompt-injection rejection. These validate the recorded source checkpoint rather than establish production accuracy or availability.
Watch the spoken product demo for a scripted student–Vox conversation and reviewed task/planning flow.
The workspace screenshot shows the development build with sample data. The spoken product demo is a scripted illustration with generated student and Vox voices and illustrated task/planning actions. Live account connections and voice behavior depend on configured providers. Undo is supported for specified calendar writes and checks the resource version before applying a change.
Current source code and supporting design records, alongside an earlier architecture snapshot.
An October 4 architecture snapshot. It predates the current LiveKit voice runtime and 5,985-test checkpoint described above.