CodeNinja has published the technical requirements for a self-improving AI system deployed at a large gas processing operator in Saudi Arabia. The system, built on CodeNinja Praxis, manages a fleet of gas compressor trains by writing operational outcomes directly into the AI's reasoning context rather than retraining model weights. This approach aims to close the feedback loop between maintenance predictions and actual inspection findings, a gap that traditionally causes reliability engineers to relearn the same judgments train by train.

Architecture: Memory Over Weights

The core innovation here is the rejection of continuous fine-tuning in favor of a governed ontology and a memory-based learning loop. The system utilizes GLM 5.3 open-weight models, self-hosted within the operator's data center in Saudi Arabia to ensure data sovereignty. When a work order closes, an outcome agent compares the initial prediction with the actual finding. This pair is stored as a memory entry that subsequent recommendations must cite, creating an auditable trail of lessons learned without altering the underlying model parameters.

Agent Roles and Human Oversight

The workflow is divided into three named agents: a condition agent for risk ranking, an interval agent for inspection scheduling, and an outcome agent for lesson drafting. Each agent's output requires explicit human approval from a reliability engineer, maintenance superintendent, or maintenance planner. This 'read first' rule ensures that no agent can autonomously change a maintenance plan or protection setting, keeping human operators firmly in control of critical infrastructure decisions.

Operational Impact and Metrics

The system is designed to address the high cost of unplanned trips and unnecessary inspections. A 2026 maintenance case study cited in the requirements showed that moving from fixed maker intervals to an extend-and-inspect approach extended inspection intervals by 20 to 34 percent on one frame family. The first phase of deployment covers one plant's gas turbine-driven compressor trains for twelve months, measuring success against baselines for unplanned trips, deferred inspections, and prediction hit rates on closed work orders.

Key Takeaways

  • Self-improvement is achieved through context injection and memory records, not model retraining.
  • Data sovereignty is enforced by using GLM 5.3 open weights hosted locally in Saudi Arabia.
  • Human-in-the-loop approvals are mandatory for all maintenance plan changes.
  • The system integrates with existing plant historians, vibration systems, and lube oil labs.

The Bottom Line

This design proves that for critical infrastructure, auditable context loops beat opaque model retraining every time. By keeping the learning signal in human-readable memory entries, CodeNinja Praxis offers a path to reliability that is both sovereign and strictly accountable.