I designed and deployed a production-grade, self-correcting AI content system that generates, evaluates, and refines LinkedIn posts using multi-LLM orchestration and human-in-the-loop workflows.
The Problem
Most AI content tools rely on one-shot generation. The result is predictable: robotic tone, generic advice, and posts that scream “AI-written.”
For individuals and companies that rely on LinkedIn for visibility, authority, and lead generation, this is a serious problem. Content must feel human, intentional, and aligned with brand standards — not just syntactically correct.
The challenge: build an AI system that doesn’t just generate content, but evaluates and improves it automatically, while keeping humans in control.
The Solution
Redraft is a self-correcting AI content engine designed to operate like an editorial system rather than a text generator.
It combines:
- One LLM optimized for creative generation
- A second LLM optimized for objective evaluation
- A recursive rewrite loop
- Human approval at critical decision points
The result is an AI pipeline that continuously refines output until it meets predefined quality thresholds or is escalated for review.
System Architecture
Redraft follows a decoupled, event-driven architecture designed for automation at scale.
High-level flow:
- Content topics are provided manually or generated automatically
- GPT-4 generates an initial LinkedIn post draft
- Gemini 2.0 Flash evaluates the draft against quality criteria
- If the draft fails, structured feedback is returned
- The system triggers an automated rewrite loop
- Approved drafts are persisted to a database
- Humans review and approve via web dashboard or Telegram
- Final content is prepared for publishing workflows
- The entire workflow can be sheduled to run daily through Github actions
This design cleanly separates generation, evaluation, storage, and human oversight.
Key Engineering Decisions
Multi-LLM Orchestration
GPT-4 is used for creative reasoning and tone, while Gemini 2.0 Flash acts as an objective evaluator. This separation mirrors real editorial workflows and avoids the weaknesses of single-model systems.
Cost-efficient Recursive Evaluation Loop
Instead of regenerating blindly, the system feeds structured evaluator feedback back into the generator, improving quality iteratively and reducing prompt heaviness. After max_retries are reached, the draft is handed over to a human.
Human-in-the-Loop (HITL) Design
AI handles scale and iteration; humans retain final authority. Approval and rejection actions directly influence the rewrite pipeline.
Persistent State & Observability
All drafts, feedback, and actions are logged in Supabase (PostgreSQL), enabling traceability, analytics, and future optimization.
Core Capabilities
- Self-correcting AI content generation with evaluation loops
- Brand-aware rewriting to eliminate AI buzzwords and robotic tone
- Web dashboard for review, statistics, and system control
- Telegram bot for real-time approval or rejection
- Automated daily content generation via scheduled workflows
- Database-backed “kill switch” to pause automation instantly
Production-Grade Implementation
- Backend: Flask-based API with modular workers
- Persistence: Supabase (managed PostgreSQL)
- Deployment: Fully Dockerized with Gunicorn
- Automation: GitHub Actions for scheduled execution
- Reliability: Safe execution wrappers with retry logic and exponential backoff
- UI: Lightweight vanilla JS dashboard optimized for speed and clarity
This system was designed with real operational constraints in mind — not just demos.
Outcome & Impact
- Built an end-to-end AI content pipeline that mirrors professional editorial workflows
- Reduced manual content iteration by automating evaluation and rewriting
- Created a reusable system adaptable to personal brands, agencies, or SaaS teams
Redraft demonstrates how AI systems can move beyond novelty and into revenue-supporting infrastructure.
Live Demo
▶ View the live Redraft system on Hugging Face
Includes dashboard, evaluation loop, and approval workflow running in a production-style environment.
Skills Demonstrated
- Multi-LLM System Design
- Prompt Chaining & AI Evaluation Loops
- Human-in-the-Loop Architectures
- AI Automation & Scheduling
- Production Deployment with Docker
- Business-Aligned AI System Design
- Prompt Engineering
Planned Improvements
- User authentication and role-based access
- Post performance feedback loop (engagement-aware rewriting)
- Multi-platform expansion beyond LinkedIn
- Scoring visibility for evaluator outputs
Why This Project Matters
Redraft demonstrates how AI-generated content can be treated as a controlled system rather than an uncontrolled output — where cost-efficency, quality, consistency, and approval are built into the workflow, not handled after the fact.
Want such a system custom built for your business?
I specialize in designing and deploying production-grade AI agents that solve real operational challenges. Let's discuss how we can automate your high-stakes workflows.
Contact Me