CARDBOARD PROTOCOL / PROJECT DOSSIER
RECALL
Internet video converted into structured, searchable knowledge.
Chapter 01
The Trigger
The internet already contained the courses, but not a reliable learning system around them.
The content already existed. The missing part was a system capable of turning it into knowledge.
Open tabs
YouTube playlists Coursera
notes bookmarks
search history partially watched videosFIELD OBSERVATION
The internet already contained the courses.
What it did not contain was a reliable way to turn those courses into a learning system.
A saved video is not yet knowledge.
Chapter 02
The Hidden Assumption
The video player was being treated as the learning environment instead of a source.
ASSUMPTION
Player as destination
A player provides playback, timeline, comments and recommendations.
It does not naturally provide curriculum, searchable transcripts, prerequisites, persistent context, review questions or semantic retrieval.
The first boundary to change was simple: the video should become a source, not the final interface.
Chapter 03
The First Model
Convert one external URL into an internal learning object.
First useful flow
- 01URL
- 02Metadata
- 03Learning Space
- 04Transcript
- 05Notes
- 06Progress
- 07Search
Learning Spaces
Spaces represent a subject or learning goal: Machine Learning, Advanced React or System Design.
The first model focused on one action: paste a URL and convert the external video into an internal learning object.
Chapter 04
The Friction
A URL looked like one input but expanded into a long-running, failure-prone pipeline.
- Videos, playlists and courses
- Duplicate or missing metadata
- Unavailable captions and long audio
- Transcription cost and job duration
- Partial failures and multiple languages
- Inactive workers and indexing delays
- Context-window and curriculum-order constraints
The interface made ingestion look like one action.
The system revealed that it was a pipeline.
Chapter 05
The Breakthrough
Model learning conversion as separate asynchronous jobs with independent failure boundaries.
Ingestion pipeline
- 01Validate source
- 02Detect platform
- 03Create job
- 04Extract metadata
- 05Queue transcript
- 06Index content
- 07Summarize
- 08Rebuild curriculum
PROTOCOL
Captions-first
Existing captions are faster, cheaper and avoid unnecessary GPU work.
Whisper remains necessary because not every source contains usable captions.
Independent queues
- Ingestion
- Transcription
- AI summary
- Curriculum reconstruction
- Independent retry and scaling
- Failure isolation and observable status
Chapter 06
The System
Web capture, API state, durable data, queues and workers form one inspectable pipeline.
System boundary
- 01Browser / extension
- 02Next.js
- 03FastAPI
- 04PostgreSQL
- 05Redis queues
- 06Workers
Worker topology
ingestion yt-dlp
transcript FFmpeg + faster-whisper
summary heuristic / OpenRouter / Ollama
curriculum dependency graph + topological sort
search MeilisearchAI cannot be a single point of failure
The heuristic provider preserves baseline functionality when remote or local model providers are unavailable.
Manual curriculum reordering persists immediately without rebuilding the entire curriculum.
Chapter 07
Trade-offs
The pipeline gains specialization while accepting explicit operational complexity.
Next.js + FastAPI
The interface stays in Next.js; Python-native ingestion and transcription stay in FastAPI.
Trade-off: two runtimes and more deployment complexity.
PostgreSQL + Meilisearch
PostgreSQL remains the source of truth; Meilisearch serves retrieval.
Trade-off: indexing consistency must be managed.
RQ + Redis + extension
Explicit jobs and in-context capture reduce user friction.
Trade-off: worker supervision, retry, authentication and configuration become operational concerns.
Chapter 08
Current Limits
Quality and availability depend on source media, workers, indexes and configured providers.
CURRENT LIMIT
- Transcription depends on audio quality and model size
- Larger Whisper models require stronger hardware
- Curriculum quality depends on metadata and provider
- Heuristic AI is reliable but lower quality
- Workers must run for asynchronous actions
- Search waits for indexing completion
- External platforms may change
- Job status currently uses polling
- Some providers require API keys
Chapter 09
What I Would Change
The next iteration should make workflows more durable, attributable and resource-aware.
NEXT EXTRACTION
- Event-driven status instead of polling
- Durable and resumable workflows
- Transcript confidence and source provenance
- Summary citations linked to timestamps
- Semantic and hybrid search
- Spaced repetition and knowledge graph
- Stronger observability
- Resource-aware transcription scheduling
- User-controlled models and offline study exports
Chapter 10
Technical File
The tools and boundaries behind the learning conversion pipeline.
VERIFIED
web: Next.js + Zustand + browser extension
api: FastAPI + Pydantic + JWT
data: PostgreSQL + SQLAlchemy
queues: Redis + RQ
media: yt-dlp + FFmpeg + faster-whisper
search: Meilisearch
providers: heuristic / OpenRouter / OllamaRecall began by saving videos. The system only became coherent when the video stopped being the destination and became raw material for a learning structure.
Back cover
The video became raw material for a learning structure.