Skip to content

extraction.md

Latest commit

 

History

History
49 lines (38 loc) · 2.36 KB

extraction.md

File metadata and controls

49 lines (38 loc) · 2.36 KB

MindOS v2

project name + description

  • mindos-v2: Rebuilt YouTube-to-email course platform — FastAPI backend, SQLite, deployed at mindos.fly.dev

who it's for

  • Self-learners aged 18–35 who save YouTube playlists but never finish them

current status

  • in progress — deployed at mindos.fly.dev

what was actually built

  • FastAPI backend with SSE streaming for live progress updates during transcript processing
  • SQLite database with tables: Users, Courses, Queue, Content, Sessions
  • APScheduler with three jobs: send_due_emails (every 5 min), generate_upcoming (every 1 hr), sync_to_sheets (every 1 hr)
  • Worker pool for background transcript processing
  • Spaced repetition: auto-generates review emails on days 3, 7, and 30 after a video
  • Cross-video synthesis email every 5 videos
  • Admin HTML dashboard (55KB) at /admin
  • Docker + Fly.io deployment (fly.toml present)
  • Transcript pipeline: youtube-transcript-api → yt-dlp fallback
  • Endpoints: POST /api/enroll, GET /api/job/{id}, GET /api/job/{id}/stream (SSE), GET /api/courses, DELETE /api/courses/{id}, GET /api/cron
  • Frontend: single-page index.html + admin.html

why it was built

  • v1 failed due to Streamlit timeouts, Google Sheets payload limits, no async, and JSON fragility from cheap models; v2 rebuilt everything on a proper async stack

blockers or reasons shelved

  • n/a — not shelved; actively deployed and in progress

wins or progress moments

  • SSE streaming gives users live progress feedback during enrollment (no more timeout UX)
  • Spaced repetition and synthesis emails are a step beyond v1 — adds actual pedagogical value
  • Docker + Fly.io deployment makes it a real hosted product

pain points

  • OpenRouter (Gemini Flash / Qwen) JSON fragility still a recurring issue — planned migration to Claude API with Pydantic/instructor
  • Keeping APScheduler jobs in sync with the worker pool across restarts

where claude api / ai was used or planned

  • Currently: OpenRouter (Gemini Flash / Qwen) for transcript → HTML email generation
  • Planned: migrate to Claude API for better JSON reliability using instructor + Pydantic

what would've helped

  • Claude API access earlier — would have avoided the JSON fragility issues from cheap models
  • Persistent storage on Fly.io (SQLite doesn't survive deploys without a volume)

metrics or traction

  • none yet — deployed but not publicly launched