Autonomous Content AI Agent, own project
Autonomous Content AI Agent
Professional content 24/7: blog + social + newsletter on autopilot
The figures
- Articles generated37 personal brand + 321 directory
- 358+
- Average quality scoreout of 100 points
- 95.6
- Social posts publishedacross 6 platforms
- 1,621+
The problem
Maintaining constant online presence is exhausting. Blog, social media, newsletter... you either work on content or on your business. And scaling across multiple sites makes it impossible.
What I did
Two coordinated systems: one for personal brand (4 pipelines, 7 data sources, bilingual generation) and another for gestorias directory (14 AI agents, 8 rotating formats, Jaccard deduplication). Both autonomous, both with quality gates.
The result
From 32 articles in November (launch) to 358+ combined across both systems. Average quality score of 95.6/100. Articles below 60 are rejected, 60-79 automatically refined. At full capacity, both systems generate 4,380+ articles per year across 6 social platforms.
The same, for your business
- A clinic that shows up first when people search «dentist near me».
- A garage that gets calls from Maps.
- A law firm that comes up when someone asks an AI assistant.
Technical details
Architecture
- Average quality score of 95.6/100 with automatic 8-point checklist
- Zero-API deduplication: Jaccard keyword similarity algorithm (45% threshold)
- Anthropic prompt caching with up to 80% input token savings
- 14 coordinated AI agents for the gestorias directory system
- 11 Supabase Edge Functions for personal brand + bilingual ES/EN generation
- 8 rotating article formats to avoid repetition
- 2,084 topics queued - over 500 days of planned content
Technical decisions
- Claude Haiku 4.5 for all 14 directory agents: 90% of Sonnet capability at a fraction of the cost
- Prompt caching on repeated system prompts: 80% input token savings
- Zero-API Jaccard deduplication: keyword similarity without embedding cost
- Single-call social generation: 5 platform posts in one Claude call instead of 5 separate calls
- Sanity CMS for structured blog content with Portable Text
- Supabase Edge Functions for personal brand: minimal latency from the edge
How it was built
Two autonomous systems merged into one showcase. The first (personal brand) uses 11 Supabase Edge Functions with 5 daily pipelines: Fetch Sources, Process Content, Publish Fallback, Newsletter and Statistics. Generates bilingual content from 7 data sources with 8-layer deduplication. The second (directory) operates 14 AI agents generating SEO articles, evaluating with LLM-as-Judge (0-100 rubric, 70 threshold) and publishing to Sanity with Portable Text. Both systems run without human intervention.
Tech stack
- Claude Sonnet 4.6
- Claude Haiku 4.5
- Supabase Edge Functions
- Sanity CMS
- Late API
- pgvector
- Vercel Cron
- Resend
- Gmail API
Features
- Wake up to new content published - without writing a word yourself
- Keep your audience engaged - posts on 6 social networks effortlessly
- Weekly newsletter that writes itself - your subscribers get constant value
- Guaranteed quality content - only content passing quality gates gets published
- Smart deduplication - Jaccard algorithm prevents repeated content without API cost