How Red Hat’s aircraft-lease quickstart hands work across 10 AI agents

Red Hat’s aircraft-lease quickstart is useful beyond aviation because it makes the handoffs inside a multi-agent application unusually explicit. The engineering post and deployment guide package the workflow as a runnable OpenShift AI example rather than a diagram.

Follow the 10 handoffs

The pipeline begins with Contract Intake, which validates an upload, classifies the document and selects its processing path. Term Extractor turns contract text into dates, financial figures, aircraft identifiers, parties and special clauses. Obligation Mapper then assigns duties, deadlines and responsible parties.

The middle of the chain connects contract language to operating data. Utilization Reconciler compares contracted flight hours and cycles with maintenance data. Reserve Calculator reads the relevant clauses but performs reserve arithmetic deterministically. Variance Detector identifies differences between the agreement and actual payments, schedules or condition reports, while Return Readiness turns redelivery requirements into a gap analysis and cost estimate.

The final stages make the results reviewable. Evidence Pack links findings to source clauses. Decision Support assembles a return, extension or buyout analysis. Escalation routes material variances, compliance flags and deadlines to people with the evidence attached. That last handoff matters: the reference pattern does not present model output as the final authority.

Know what the install requires

The quickstart targets OpenShift 4.19 or later and OpenShift AI 3.4 or later. Its documented baseline adds three workers with 8 CPUs and 32GB of memory each; one GPU worker is recommended for Granite inference. The chart requests a 5Gi persistent volume and requires either cluster-admin access or admin rights in the target namespace.

A Helm install deploys the Next.js frontend, FastAPI backend, Python worker, PostgreSQL and Redis, plus vLLM, Llama Stack and IBM Granite 3.3 2B Instruct through chart dependencies. LangGraph sequences the agents, and Docling parses PDFs with a PyMuPDF fallback. The guide includes a Helm smoke test, sample contracts across 10 document types and an Argo CD example for teams that want a GitOps-managed installation.

Treat the limits as design input

GPU runs take roughly a minute per document in Red Hat’s documented walkthrough. CPU inference is supported, but individual model calls take about 40 seconds and a full lease can take 20–30 minutes depending on document size and capacity. Demo mode is faster only because it skips the worker and model, writes synthetic extraction data and produces no live agent progress or audit events.

The default 2B model is deliberately small enough for modest hardware, and the troubleshooting guide warns that it can misread dates and figures. Results should be checked against extracted terms or produced with a larger model. The in-cluster PostgreSQL and Redis are single-replica quickstart defaults; external services are supported for a more durable setup. Red Hat also says the content has not been tested on every supported configuration, while the Helm configuration is Apache 2.0 and the Codvo application images are proprietary.

The reusable lesson is architectural: separate probabilistic extraction from deterministic calculations, preserve clause-level evidence, and make the final escalation a human decision point. The quickstart is a deployable way to inspect those boundaries—not a production-ready contract-review system.