Draw the boundary first
An air-gapped workbench begins with an operational boundary, not a hosting label. Identify what must run inside the environment and how people will operate it. A proposed path should not require an external inference endpoint, a remote license check, or a support service that the environment cannot reach.
This article describes a planning method, not a published customer deployment. Superbloom can be scoped for on-premises and air-gapped environments. The actual configuration depends on the team’s models, tools, compute, and access policies; those dependencies deserve explicit review before an engagement begins.
Inventory a real workflow
Choose one recurring analysis or test process and list the artifacts and tools it uses. Follow the path from input preparation through execution and review. Include scripts, material data, supporting libraries, licenses, and the locations where evidence is stored. Small utility dependencies are easy to miss when attention stays on the main solver.
For each step, ask whether it can run within the proposed boundary. An installed application may still rely on remote activation, updates, or authentication. Confirm the permitted operating mode with the responsible tool owner. Treat unverified dependencies as scoping questions, not assumptions that will resolve themselves.
Match models and compute to the task
Take an inventory of the available machines, memory, accelerators, storage, and job scheduling arrangements. Solver work and AI inference may compete for different resources. Decide whether those workloads can share a machine or need separation, then evaluate the candidate setup on representative tasks.
Local model availability is part of that assessment. Review the model’s terms, the inference runtime, and the process for transferring approved weights and dependencies into the environment. Model size alone is not an acceptance criterion: the relevant question is whether the complete workflow performs adequately under the team’s engineering checks.
Plan movement, maintenance, and recovery
Agree how approved artifacts enter and leave the environment. Define the transfer path, the review responsibility, and how versions are recorded. The same care applies to software updates and diagnostic information used for support. An update process that depends on unrestricted downloads is not a complete air-gapped plan.
Establish who can operate the workbench, modify workflows, and approve engineering decisions. Decide how logs and evidence are backed up and restored. Support arrangements should state what can be inspected locally, what may be shared, and who authorizes that sharing. Specific controls and responsibilities belong in the engagement agreement.
Prove the path before expanding it
Run one bounded workflow from approved inputs to retained evidence and an engineering review. Check the path with external network access unavailable, including restart and recovery behavior. Measure the same task scope as the baseline and record the effort required to operate and maintain the setup.
Expand only after the environment and method meet the agreed criteria. The first useful result is an operationally credible workflow that the team understands, with remaining dependencies and limitations made visible.
- Compute and storage inventory
- Approved models, inference runtime, and transfer process
- Solver interfaces, dependencies, and offline licensing
- Access, artifact movement, and evidence storage
- Updates, backups, recovery, and support responsibilities
- Baseline task, acceptance criteria, and engineer review