CrossnoKaye's ATLAS cuts plant energy 21% across 13,000 assets

CrossnoKaye's approach differs from the common pattern of bolting an external AI agent onto an industrial stack. Plant data and control logic are typically spread across separate systems from different vendors—PLCs, DCS, historians, SCADA, and standalone tools—each with its own naming conventions, while the actual control sequences and safety interlocks are locked inside proprietary controller code. Distributing credentials across the plant network is another hurdle. ATLAS instead defines the plant in software from the edge up, with an on-premises control panel running a local Kubernetes cluster and streaming telemetry to AWS through Amazon VPC Lattice and Amazon MSK. This architecture gives AURA native access to the full operational state within a single security boundary.

AURA's operational modes include Chats for interactive engineering conversations, Tasks for repeatable parameterized investigations, and Routines for automated background watchers. For Chats, the Sandbox Orchestrator binds one warm Amazon Bedrock AgentCore session with a 15-minute sliding TTL; for Tasks and Routines, it provisions a fresh microVM per run and stops the session immediately, preventing cross-run contamination and idle charges. CrossnoKaye reports that combining software-defined edge control, continuous cloud optimization, and AURA has reduced batch drying-time standard deviation by 34 percent and identified more than 50,000 tons per year of recoverable production gap in quarry operations.

Three lessons stand out from the production deployment. First, unifying the plant model and control logic before adding AI lets an agent distinguish a slipping downstream recirculating belt from a secondary crusher fault. Second, embedding AURA into the daily shift rhythm—background Routines catching anomalies overnight and delivering a cited Brief to the engineer's inbox—drives adoption better than a separate tool requiring a blank prompt. Third, enforcing data, coverage, and safety contracts in code rather than prompts catches probabilistic model errors before they reach an operator or a piece of equipment.