project creatortrack social voice profile pipeline 20260715T174742
name: CreatorTrack social data layer + AI voice profile pipeline (Tasks 13-18, M1-M4) description: Multi-task build on feat/social-data-layer-and-scripts, ws85: transcript ingest, cross-post matching, retention-weighted voice profile, script generation next type: project
Active multi-task build for CreatorTrack's social/creator-tools product, branch feat/social-data-layer-and-scripts, workspace id 85 (Gus the Bass channel, 515 videos across YouTube/TikTok/Instagram/Facebook).
As of 2026-07-15:
- Task 13: transcript + structure extraction complete (515 transcripts, hook taxonomy, words-per-second, sentence length).
- M1-M3 (Meta lane, cross-post matcher): complete. Fixed a bug where a 21-day published-date window excluded true cross-platform reposts (Gus reposts up to ~3 years / 1168 days later) — see [[project_gus_crosspost_repost_window]]. Added app.social_crosspost_exclusions table + --reject CLI for human judgment calls on ambiguous pairs. 175 auto-linked, 650 ambiguous awaiting human review, 1273 no-match.
- M4: platform filter chips + cross-post nav UI, building.
- Task 14: retention-weighted voice profile (app.voice_profiles, ws85 "Default", id=3) synthesized and reviewed; found and is fixing a weightedPercentile function that didn't actually apply retention weights (structural bug, small numeric impact on this corpus). Design principle: voice card is hand-editable in the Scripts app UI, and any "regenerate" action must never silently overwrite a human's hand-edits.
- Task 15 (next): embeddings + retrieval + generation endpoint — first task that writes an actual script in Gus's voice.
- Result files land at comms/results/creatortrack-social-*.md and comms/results/creatortrack-social-data-layer-and-scripts-app-task-*.md.
Why: this is Justin's creator-tools product idea ([[project_creator_tools_product_idea]]) taking concrete shape — turning a creator's own catalog into a voice-matched script generator. How to apply: when picking this back up, check task numbering above for what's done vs. pending before re-dispatching workers; don't re-run Task 13/14 extraction.
[auto-memory session 60d132de-4659-485c-8044-427af0ccce7f, confidence 0.72, mode staged]