Creator Cortex — YouTube Title Analyzer & Strategy Platform

Internal YouTube strategy product — a FastAPI backend and React 19 + Vite studio for serious creators: score titles, generate alternatives with RAG over real high-performing outliers, mine competitors and patterns, connect multiple channels via OAuth to pull YouTube Analytics, and close the loop so feedback and performance shape the next generations.

The problem

Guessing titles from gut feel doesn’t scale. Creators need data (what’s worked before, on their channel and elsewhere), fast iteration, and one place that ties scoring, generation, outliers, and post-publish outcomes together—without juggling spreadsheets and disconnected tools.

What I shipped

• Title scoring — Heuristic + structured scoring (length, clarity, curiosity, specificity, patterns) with batch and “health” style endpoints the UI can call on demand.
• LLM title generation (RAG) — Central gateway to OpenRouter with tiered models (e.g. Claude for heavy generation, smaller models for classification). pgvector retrieval over embedded outlier titles so prompts are grounded in real winners, not generic fluff.
• Outlier & competitor mining — Scheduled jobs (APScheduler) refresh channels, scan for outperformers, backfill embeddings, and support formula / pattern discovery the UI can browse.
• YouTube OAuth + Analytics — Multi-channel connect, encrypted token storage, pulls real metrics (CTR, impressions, retention signals, etc.) to power dashboards and feedback-aware prompt blocks.
• Feedback loop — Tracks generations, picks, edits, and post-publish snapshots so later prompts can emphasize what actually worked for this creator.
• Frontend studio — Multi-tab app (scorer, generator, outliers, channels, insights, thumbnails playbook, etc.) consuming the large generated API surface.

Architecture (at a glance)

• Modular FastAPI — Feature-oriented packages (routes, services, repositories) so scoring, RAG, OAuth, schedulers, and streaming traces stay separated.
• Async SQLAlchemy 2 — Async Postgres; Supabase hosts the database; vector search lives alongside relational models.
• SSE traces — Long-running generations can stream structured progress to the UI for transparency.

Stack

Python 3.12, FastAPI, Pydantic, SQLAlchemy 2 async, Alembic, PostgreSQL + pgvector (Supabase), React 19, Vite, OpenRouter (Claude / Gemini / GPT-4o-mini), YouTube Data + Analytics APIs, APScheduler, Docker, Render (internal deploy).

Result

A single internal product where creators move from “idea / transcript” to ranked, judged title options with retrieval and channel history in the loop—plus real Analytics-backed learning over time. Not a public SaaS: built and run for internal use; no public URL or demo link.