Turn your expertise into a trusted AI teaching presence—without surrendering the work that makes you valuable.
Alice helps educators and creators turn years of hard-earned knowledge into a trusted digital presence that can teach, answer questions, and create opportunities around the clock—without surrendering their private source material to the cloud. Creators retain ownership and control of their intellectual property, giving them the confidence to grow their authority, reach more people, and build a business from their expertise while their original work remains privately theirs.
Alice watches one folder of creator-owned videos. It transcribes selected recordings locally, writes portable Markdown, removes likely personal information on the Mac, and asks the creator to approve the exact safe copy before any knowledge can leave the device.
Approved text becomes private retrieval-augmented knowledge. The creator can test answers with timestamp citations, correct the source when something is wrong, and eventually publish a biography-and-links page with an optional AI teaching assistant.
“Training” in Alice means retrieval-augmented generation (RAG), not fine-tuning a model on the creator’s raw videos.
- Choose a folder containing
.mp4,.mov, or.m4vteaching videos. - Include or exclude recordings and resolve possible duplicates.
- Transcribe locally with resumable checkpoints for videos up to three hours.
- Review simple Markdown with timestamps, anonymous speaker labels, search, and save.
- Remove personal information locally and inspect the separate safe-copy candidate.
- Approve the exact safe copy before private cloud indexing.
- Test Alice privately with grounded answers, timestamp citations, and honest refusals.
- Prepare a Creator Home combining biography, featured content, useful links, email capture, and optional chat.
flowchart LR
A[Teaching videos<br/>on the Mac] --> B[Local<br/>transcription]
B --> C[Creator-owned<br/>Markdown]
C --> D[Local privacy<br/>review]
D -->|Approved safe text only| E[Private<br/>RAG index]
E --> F[Cited<br/>AI answers]
F --> G[Creator Home<br/>and link in bio]
Alice shows what it replaced, keeps the original local, and requires an explicit creator approval.
Private answers cite the approved teaching source by timestamp. If the knowledge does not support an answer, Alice refuses instead of guessing.
Chat is one optional module inside a broader creator page with a biography, featured destination, grouped links, social profiles, and consent-based email capture.
- Folder-first workflow — point Alice at one library instead of importing videos individually
- On-device transcription — AVFoundation audio extraction and FluidAudio/Parakeet speech recognition run on the Mac
- Background progress — visible queue states, file progress, pause controls, and resumable checkpoints
- Creator-owned output — timestamped Markdown is written beneath
Alice Transcripts/inside the selected folder - Simple transcript review — one searchable editor rather than thousands of line-by-line controls
- Local privacy pass — likely names, emails, phone numbers, addresses, account identifiers, and secrets are replaced before approval
- Hard upload boundary — the sync module accepts only a typed
ApprovedSanitizedDocument - Grounded private preview — answers cite approved sources or return an honest refusal
- Versioned publishing foundation — private knowledge and public releases remain separate choices
- Creator page and leads — biography, links, optional AI chat, explicit interests, email confirmation, and encrypted lead sync
- Accessible themes — native light/dark appearance, non-color state labels, visible focus, and reduced-motion support
Alice is designed so raw creator media has no upload path.
Alice never uploads from the watched teaching folder:
- source video or extracted audio
- the raw machine transcript
- original filenames or absolute local paths
- encrypted transcription checkpoints
- detected private identifiers
Alice may upload only after explicit approval:
- sanitized teaching text
- anonymous timestamp locators used for citations
- normalized speaker roles
- a creator-approved safe title
- random source/version identifiers and a sanitized-content hash
The raw Markdown remains local. If the creator chooses a Dropbox, iCloud, or similar synchronized folder, that storage provider may independently synchronize the Markdown file.
Read the deeper background-sync architecture and V1 implementation evidence.
The demo uses isolated synthetic teaching material and a local preview sink. It does not upload data or create a public page.
- macOS 14+
- Apple Silicon Mac
- Xcode 26+
- XcodeGen
git clone https://github.com/productdave/Alice.git
cd Alice
./scripts/bootstrap-macos.sh
./scripts/run-macos-demo.shTo run Alice against a folder you choose:
./scripts/run-macos.shThe first real transcription downloads the revision-pinned English Parakeet model. The bundled speaker-diarization components and downloaded speech model are documented in ThirdPartyNotices.md.
./scripts/test-all.shThe suite covers the native macOS app and privacy boundary, cloud ingestion and retrieval, public web contracts, accessibility checks, and the creator-to-public-release golden path. The real on-device ASR fixture is opt-in:
./scripts/test-real-asr.sh| Layer | Technology |
|---|---|
| macOS desktop | SwiftUI, AppKit, AVFoundation, sandboxed XPC helper |
| Local transcription | FluidAudio 0.12.6, Parakeet TDT v2 Core ML |
| Speaker diarization | Revision-pinned FluidInference Core ML models |
| Local storage | Encrypted SQLite/WAL state and security-scoped bookmarks |
| Privacy | Deterministic identifier detection plus on-device name/place detection |
| Cloud API and workers | Node.js, TypeScript, PostgreSQL, encrypted documents, durable queues |
| Retrieval | Hybrid vector/keyword search with source-version authorization |
| Public creator page | TypeScript, Vite, responsive accessible UI |
| Testing | XCTest, Node test runner, Vitest, contract and end-to-end checks |
.
├── apps/
│ ├── macos/ # native desktop app, XPC helper, and tests
│ ├── cloud/ # private ingestion, retrieval, publishing, and leads
│ └── web/ # public Creator Home and chat
├── packages/
│ ├── contracts/ # strict cross-language safe-document contract
│ └── model-manifests/ # pinned local-model manifests
├── docs/ # product, privacy, architecture, UX, and test plans
├── scripts/ # build, test, packaging, and pilot verification
├── assets/ # README cover and product screenshots
└── project.yml # XcodeGen project definition
Alice is a working developer preview and private-pilot foundation—not yet a hosted production service.
The repository includes the native folder/transcription/review/privacy journey, typed safe-copy sync, private cited chat, Creator Home editor, cloud foundation, public web experience, and automated tests. A public production launch still requires production identity/cloud configuration, a signed and notarized macOS build, and clean-machine pilot evidence.
Granola import, direct .md/.txt source ingestion, slide parsing, multilingual models, billing, CRM integrations, and local chat models are intentionally deferred.
- Creator persona
- End-to-end creator journey
- Current implementation plan
- Experience design
- Local transcription decision
- Test plan
- Preserved visual archive
Alice uses FluidAudio and revision-pinned Core ML speech and diarization models. Licenses, revisions, changes, and required attribution are recorded in ThirdPartyNotices.md.