PROJECT 03

Completely Autonomous Social Pipeline

A 3D printing brand needed to build an audience without a marketing team. Two posts a day, every day: AI-written, AI-imaged, automatically published, engagement-tracked, and top performers flagged for advertising. The whole loop runs itself.

AGENT FRAMEWORKSLLM INFERENCEIMAGE GENERATIONMETRICS TRACKINGSOCIAL PUBLISHING

Problem

A 3D printing brand trying to build a following in a crowded niche. The content the audience wants, maker tips, print showcases, technical breakdowns, design inspiration, takes real time to produce. Research a topic, write a caption with the right voice and hashtags, generate or source an image, post at the right time, then do it again tomorrow.

At the pace that actually builds an audience, you are looking at two posts a day. That is sixty posts a month. No small business owner can sustain that alongside running the actual business. Hire a social media manager and you have a full-time salary for a channel that might not pay off for a year.

The existing process had a human writing every caption, generating every image, and manually checking engagement. It was unsustainable at any real cadence, and inconsistent posting is worse than no posting.

What I built

A fully autonomous social content pipeline that runs twice a day without human intervention:

  • A structured content calendar that rotates through twelve content types in a fixed grid cycle: product showcases, maker tips, technical deep-dives, art prints, design inspiration. Every post has a strategic role and the feed always has variety.
  • Topic deduplication that injects the last fifteen posts of each content type back into the generation prompt as negative examples, so the system never repeats an angle or recycles a caption
  • AI copywriting and image generation through purpose-chosen models, producing captions with hashtags and original visuals. No stock photos, no templates.
  • Simultaneous publishing to Instagram and Facebook with platform-specific formatting, logging post IDs and metadata back to the content database automatically
  • Every post logged to a central Airtable database on publish: caption, image prompt, content type, post ID, platform. Every downstream workflow has a single source of truth to read from.
  • A daily engagement workflow that pulls likes, comments, and reach from both platform APIs every morning and writes the numbers back to each post record. No dashboards. The system updates itself.
  • An ad optimization workflow that ranks every post by engagement score, identifies the top performers, and flags them in the database for paid promotion. The business gets a prioritized list of what’s worth putting advertising budget behind, automatically.

The pipeline runs in three stages, in sequence, every day. Publish at 10am and 3pm. Metrics update every morning. Top posts surface for advertising without anyone pulling a report. The full loop: create, publish, measure, rank, flag. Runs itself.

What I learned

The first engagement tracker used a clever branching pattern and quietly lost posts on loop-back. Rewriting it as a linear flow eliminated the bug and made the whole workflow easier to reason about. Deduplication was the other unlock. Without it, any autonomous system eventually starts sounding like itself on repeat. Feeding recent posts back as negative examples solved it completely. Linear pipelines and explicit memory. Every time.

Technical architecture

CONTENT CALENDAR
A structured 12-post grid cycle rotates content types in a fixed pattern: product showcases, maker tips, technical deep-dives, design inspiration, art prints. Every post has a role and a slot. The feed looks curated because it is.
DEDUPLICATION
Recent posts of each type get injected back into the prompt as negative examples before each generation run. The system knows what it already said and refuses to repeat itself. No recycled captions, no retreaded angles.
GENERATION
Copywriting and image generation run through purpose-picked models: a fast language model for captions, a higher-quality image model for visuals. Outputs land in a versioned asset store before publishing.
MULTI-PLATFORM PUBLISHING
Each run publishes to Instagram and Facebook simultaneously, handles platform-specific formatting, and logs post IDs back to the content database for tracking.
CONTENT DATABASE
Every post is logged to Airtable on publish: post ID, caption, image prompt, content type, platform, timestamp. The database is the source of truth for everything downstream. Deduplication, metrics, and ad decisions all read from it.
ENGAGEMENT TRACKING
A daily workflow pulls likes, comments, and reach from both platform APIs, writes them back to each post record in Airtable, and surfaces a top-performer summary to chat every morning. No dashboards to check. The system tells you what worked.
AD OPTIMIZATION
A third workflow ranks posts by engagement score, identifies the top performers above a threshold, and flags them in the database for paid promotion. The business sees a prioritized list of what's worth putting budget behind, without pulling a single report.
KEY NUMBERS
2x daily
Auto-publish
3 workflows
In sequence
Zero
Manual Touch
Closed loop
Publish to Advertise