Personal ProjectFeatured

Obsidian-Agent-Post: Voice-Consistent AI Content Pipeline

System Architect & Builder|

4 months

90% (pre-checkpoints: 7/10 flagged, post: 1/10)

Generic AI Slop Reduction

69.1 → 75+ avg (6-dimension rubric)

CQS Quality Score

Manual 4h → Automated 90sec (96% time savings)

Research Time

31 voice checkpoints, 89 test cases validating quality

Content Consistency

The Challenge

Creating consistent, high-quality content at scale requires maintaining voice, avoiding generic AI slop, and validating quality before publishing. Manual editing is slow; unchecked LLM output lacks personality. Needed a system with 31 voice checkpoints, 6-dimension quality scoring (CQS), and Karpathy-style validation loop — all while generating blog posts and LinkedIn content in a consistent personal voice.

The Approach

Built multi-layered content pipeline: (1) Multi-source research agent querying 10 parallel sources (HackerNews, Reddit, ArXiv, GitHub, YouTube, RSS, Obsidian vault) with semantic deduplication, (2) CQS scoring system with 6 dimensions (Clarity, Quality, Substance, Voice, Structure, Impact) and subtype-conditional formulas, (3) 31 voice checkpoints validating tone, structure, and authenticity, (4) Karpathy loop for iterative refinement, (5) 32-item pre-publishing checklist. Tech: FastAPI + SQLModel + PostgreSQL + Next.js 14 landing page.

Key Learnings

Voice consistency requires explicit checkpoints — 31 rules prevent 90% of generic AI slop
Multi-dimensional scoring (CQS) catches quality issues LLMs miss in self-review
FAILURE: Initial single-pass generation produced inconsistent quality (69.1 avg CQS). Added Karpathy loop with 3-pass validation to reach 75+.
The content pipeline itself is a showcase-worthy product — document your own systems