Product · AI engineering
AI Social Media Engine
An autonomous pipeline that drafts, reviews and publishes across platforms — no scheduling, no context-switching, no human in the hot path.
Role
Engineer · Architect
Period
2026 — present
Source
Private product
01 — Context
Holding a consistent, high-quality presence across several platforms costs an enormous amount of context-switching and manual labour. Scheduling tools move the work rather than removing it — somebody still has to write every post, in the right voice, for the right platform.
The target was a system that understands the brand voice well enough to be trusted unattended, drafts contextual content, reviews its own output, and publishes natively — not a wrapper that queues things a human already wrote.
02 — Architecture
Three agents in sequence rather than one prompt: one writes, one critiques against the brand guidelines, one formats for the target platform. Queues sit between every stage, because LLM latency is unpredictable and platform rate limits are not negotiable.
Inputs
Brand guidelines
voice, do-not-say
Trending topics
ingested signals
Dashboard
React, real-time
Agent pipeline
Draft agent
Critique agent
Format agent
Message queues
Platforms
X / Twitter
native publish
native publish
native publish
Nothing calls an LLM or a platform API synchronously — every hop is a queued step that can fail and retry on its own.
03 — Decisions & trade-offs
Three narrow agents over one capable prompt
Draft, critique, format — each with one job. Cost: three model calls per post instead of one, so latency and spend both rise. Benefit: the critique step is the only reason the output is safe to publish unattended, and each stage can be evaluated and replaced on its own.
Queues between every stage, nothing synchronous
Serverless functions pulling from message queues. Cost: no simple request/response path to debug, and end-to-end tracing has to be built deliberately. Benefit: an LLM timeout or a platform rate limit retries one step instead of failing a whole post.
Native platform APIs rather than a publishing aggregator
Direct integration with each platform. Cost: three API surfaces, three auth models, three sets of breaking changes to track. Benefit: native formatting and real analytics back, neither of which survives an aggregator.
04 — Impact
- Runs three platforms off one pipeline — X, LinkedIn and Instagram.
- Removed manual scheduling from the loop entirely.
- Held a consistent brand voice across platforms with different formats.
- Survived rate limits and LLM outages without losing queued work.
Automating something similar?
The hard part is never the model call — it is the retry semantics, the rate limits and deciding what the system is allowed to publish on its own. Happy to compare notes.