AI in production.
Assistants, agents and training on infrastructure we operate. Evaluation included.
Languagemodelsonlivetelemetry.Agentsthatdorealworkacrossservers.Controlpoliciestrainedondigitaltwinsbeforetheyreachtheedge.Everychangerunsagainstafixedevaluationsetbeforeitships.Ourowncompanyrunsthisway.
LLM assistants over live data
Answers from a live MQTT or database feed, not a static corpus. In production for an infrastructure operator since 2026. Self-hosted or API models.
Agents via MCP
Multi-agent systems on our own MCP server across several hosts. Code, deploy, test, monitor. The same server runs Blackcycle.
Training on digital twins
Multi-agent reinforcement learning on SUMO twins with RLlib and PettingZoo. The trained policy moves to the edge node by changing one parameter. Preprint and dataset on Zenodo.
RAG and vector search
PostgreSQL with pgvector by default, Qdrant at scale. Retrieval quality measured before the model answers.
Frequently Asked Questions
Both. Open models on our own GPUs when data cannot leave the perimeter. Claude or similar APIs when quality or context matters more. Same evaluation harness for both.
A fixed set of cases with a designed correct answer, run N times per prompt version. Accuracy, p50 and p95 latency, prompt-injection resistance. No regression, no ship.
One tool interface. Agents discover and call ERP, CRM, databases, CI/CD and infrastructure through one protocol. Permissions managed in one place.
A control policy has to try bad actions to learn. You cannot do that on a live junction. The twin runs thousands of episodes. Real data calibrates it.