AMP launches AI agent Ray for renewable energy monitoring
AMP is deploying an AI agent named Ray to monitor renewable energy portfolios in Sub-Saharan Africa, using code execution to triage alerts and quantify financial impact. The company has also standardized its engineering workflow on Slack-integrated agents to improve auditability and collaboration.
AMP, a 13-14 person renewables monitoring company focused on Sub-Saharan Africa, is launching an externally facing AI agent named Ray. Ray accesses site telemetry, alert logs, and production data via an MCP server to perform portfolio analysis. It triages alerts by quantifying impact on performance ratio and monetary value based on local tariffs, prioritizing issues across hundreds or thousands of sites. The agent uses code-writing capabilities with libraries like DuckDB and Polars to process time-series data, overcoming LLM limitations with quantitative inputs. Pilots showed high accuracy in detecting faults like string disconnects and metadata errors. The system runs on a $100/month OpenAI subscription using GPT-6.1 Sol. AMP notes that high-quality underlying data is critical for agent performance, as models effectively flag data gaps and inconsistencies without heavy prompting. AMP has also standardized its 7-person engineering team on the Hermes agent framework, deploying agents like Sparky for triage and Coda for coding within Slack. Agents operate in Docker-based sandboxes on firewalled EC2 instances, with specific toolsets for different tasks. The workflow involves Sparky investigating alerts and creating Linear tickets, while Coda reviews the context, writes code, spins up sub-agents for review, and creates GitHub PRs linked to the original Slack thread. This setup improves auditability and attribution compared to local IDE usage. AMP co-founder Svet Bajlakov advocates for this Slack-integrated agent workflow as a superior method for maintaining transparency, context, and collaboration in AI-assisted software development. The company is evaluating various AI models for energy data tasks, finding that GPT-6.1 Sol offers a solid balance of intelligence and conciseness. The team is exploring JEPA classification models to pre-filter alerts and compact context before feeding data to LLMs. They conclude that base-level intelligence from recent models is sufficient for their use cases, reducing the need for bleeding-edge capabilities.