Self-Hosted AI Content Engine

Problem

Publishing consistently across multiple brands means either paying metered API costs that scale with usage, or manual effort that doesn't scale at all — and autonomous posting without review risks putting a bad generation on a public account.

Goal

Zero marginal cost per post: generate everything locally on owned hardware, route each piece to exactly one brand's blog and social accounts, and gate every publish behind a human approval click.

Key Features

• Per-brand marketing pipelines with strict account isolation (no cross-brand credential fallback)
• Permanent post history in DynamoDB with feedback-driven iteration that republishes to the same URL
• Per-platform content: long-form LinkedIn posts with an architecture diagram, caption-style posts with branded title cards elsewhere
• Automatic CloudFront invalidation after each brand deploy completes
• Architecture diagrams rendered from code (Python diagrams + Graphviz) so published diagrams match the real system

Tradeoffs and Design Decisions

• Local small-model generation is free but needs a validation layer — JSON mode constrains syntax, not semantics, so every structured field is checked against its schema
• Owned-GPU inference inverts the cost curve (zero marginal cost per post) at the price of managing local infrastructure instead of calling a metered API
• A human approval gate caps full autonomy by design — ten seconds of review per post versus the reputational cost of one bad autonomous publish

Challenges

• First Instagram publish failed with 'media fetch failed': the scheduler served images from localhost, which Meta's servers can't reach — fixed by pre-uploading all media to a public S3 bucket
• Small local models sometimes echo an entire enum list instead of picking one value — solved with schema validation and preferring curated values over generated ones
• Small models also collapse output length on revision tasks — flagship long-form content is drafted by a stronger model and published verbatim, bypassing the local rewrite step

Results and Lessons Learned

• Three brands live from one approval flow: blog update + Instagram + Facebook on one brand, blog post + LinkedIn + personal socials on another, structured case study on the third
• Total running cost ~$2-5/month (DynamoDB + S3); generation cost $0 on owned hardware
• Every post permanently tracked and iterable: feedback in, improved revision out, same URL republished