Skip to Main Content

Passaly

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Passaly — Demo video

Passaly Classes turns every class and meeting into memory that people and AI agents can use. Siena is 10 and lives in Brazil; her teacher Giovanna is in Canada, 11,000 km away. Class time is short, and questions come later, when the teacher is asleep. Passaly Classes captures what was said in the class and turns it into a personal tutor that answers the way Giovanna teaches. Questions the class didn't cover go back to Giovanna, and her answer becomes part of the class for the next student. Meetings work the same way: - Recap: everyone gets a summary, decisions, action items and open questions. - Tasks: they go into a task list, and any AI agent can pick them up, do the work and report back. - Assistant: it answers "what did we decide?" weeks later and can put things in your calendar. How it works - Transcript: each person's phone or browser captions its own speaker. Or a Google Meet meeting is created with transcription on, and Meet's transcript comes back by itself after the call. - Two AI options: - Llama 3.2 runs inside the Android phone and works offline after class, so no cloud AI sees the questions. - Muse Spark runs in Meta's cloud for recaps, the tutor and an assistant that uses tools. - One shared memory (ai-memory) holds transcripts, recaps, files, notes and tasks. Switching between the two AI options changes who thinks, not what is remembered. - MCP server: Claude Code, OpenCode or any other MCP app can read and search your classes and meetings, and complete their tasks. Each app gets its own token: read-only unless you allow writing, revocable, and it expires. - Private by design: no audio or video is stored, and you only see what you took part in. Built with: Llama 3.2 (on the phone, via LiteRT-LM), Meta's Muse Spark (Meta Model API), FastAPI, SQLite, Flutter, ai-memory, MCP and the Google Meet REST API. 3. Why both Llama on the phone gives privacy and offline use. Muse in the cloud gives stronger reasoning and tool use when there's a connection. Both read and write the same memory (transcripts, recaps, files, notes, tasks), which is also open to other AI agents through our MCP server. Moving between Llama and Muse changes who does the thinking, not what is remembered.

Team