An AI-powered content operations platform that automates research-to-draft pipelines while preserving authentic voice through multi-agent quality control and deterministic brand enforcement.

In chess, a "Centaur" is a human-AI team that consistently outperforms both humans and AI playing alone. Centaur applies this principle to content creation: the human provides voice, judgment, and lived experience — AI handles research, drafting, formatting, and analysis. The result is content that is authentically human at scale.

Professional thought leadership requires a punishing volume of high-quality content — researching trends, drafting posts, maintaining brand voice, optimizing for platform algorithms, tracking engagement, and iterating on strategy. Most professionals can sustain this for a few weeks before the pipeline collapses.

AI content tools solve the volume problem but create a new one: generic, detectable output that erodes credibility. Platform algorithms increasingly penalize obviously AI-generated content, and audiences can feel the difference.

Centaur is a human-in-the-loop content operations platform that sits between these extremes. It automates the research-to-draft pipeline while preserving authentic voice through three mechanisms:

1. Style fingerprinting — The system learns a creator's voice from their existing writing, capturing quantitative patterns that go beyond "write like this person"
2. Multi-agent quality control — Specialized AI agents draft, critique, and refine content in an iterative loop, each focused on a different quality dimension
3. Deterministic brand enforcement — A rule engine validates every draft against measurable constraints before it reaches human review, catching violations automatically and saving human attention for subjective quality judgment

The system never auto-posts. Every piece of content is queued for human review, editing, and approval. This is a deliberate architectural constraint — authenticity cannot be automated.

Instead of generating content in a single prompt, Centaur uses a Writer-Critic-Reviser pipeline — three specialized AI agents that iteratively refine content. The Writer drafts, the Critic evaluates against quality criteria, and the Reviser addresses specific critique points while preserving voice and intent. The loop continues until quality thresholds are met.

This approach produces measurably better output than single-pass generation because each agent has a focused role and evaluation criteria.

The system abstracts across three LLM providers through a unified interface:

• Local models (Ollama on Apple Silicon) handle research and first drafts — cost-free, private, low-latency
• Cloud APIs (Claude, Gemini) handle quality-critical tasks like style transfer — highest accuracy where it matters most
• Tasks route automatically based on requirements, with graceful fallback between providers

Content strategy isn't static. Centaur includes an analytics engine that ingests platform performance data, runs anomaly detection, and produces multi-signal assessments. Strategy and style rules evolve based on measured outcomes, not intuition. Every piece of content is tagged with strategy and style versions for controlled attribution analysis.

See ARCHITECTURE.md for the full system design, data flow diagrams, and design decision log.

Centaur is developed using a five-agent AI development ecosystem — multiple AI agents coordinated through shared activity logs, inter-agent handoff notes, and a cross-project message board. Each agent has defined domain lanes, composure conventions, and capability boundaries. This multi-agent development approach itself demonstrates the Centaur philosophy: humans direct, AI executes across parallel workstreams.

The project follows a phased development roadmap with formal product requirements for each phase, a technical decision log documenting every significant architectural choice (with rationale and alternatives considered), and a module registry tracking all components with input/output contracts.

This is a showcase repository — documentation only. Source code is maintained in a private repository.

Copyright 2026 TJ Neary. All Rights Reserved.