​

Build trust before automation

InferCrane’s direction is to become the trusted decision and operations layer between AI applications and heterogeneous inference capacity. The roadmap is ordered around user trust, not the number of integrations in a logo grid.

​

Now — prove the production core

• Complete real-provider and real-runtime qualification for the current release.
• Keep one stable endpoint identity across customer-operated workers and explicitly budgeted managed model APIs; credentials remain server-side references and new bindings begin as candidates.
• Make build, connect, update, rollback, autoscaling, serverless, and deletion boring and recoverable.
• Publish reproducible model/runtime/GPU evidence instead of unsupported performance claims.
• Tighten the five-minute path from installation to a successful request.

​

Next — broaden adoption safely

• Make development, staging, and production promotion explicit and reproducible.
• Qualify additional provider and runtime combinations through stable contracts.
• Improve private networking, workload identity, HA, backup/restore, and upgrade evidence.
• Turn Request Inspector, Doctor, alerts, and FinOps evidence into a daily operator workflow.
• Expand verified recipes and Inference Lab with measured, reproducible results.

​

Later — decision intelligence

• Replay privacy-preserving production workload shape against candidate serving plans.
• Learn provider capacity and cold-start reliability from grounded observations.
• Recommend model/runtime/GPU/provider plans with measured, modeled, and heuristic evidence clearly separated.
• Verify recommendations through benchmark, Release Guard, human approval, and post-change results.
• Explore policy-bounded automation only after recommendations repeatedly predict reality.

​

Dependency order toward end-to-end inference operations

InferCrane expands around one control loop rather than accumulating unrelated infrastructure:

1. Serve: build or connect a workload, then expose one stable endpoint.
2. Govern: admission, quotas, referenced credentials, budgets, private networking, and identity.
3. Observe: request evidence, deterministic Doctor findings, alerts, cold starts, and cost.
4. Change safely: immutable candidates, replay or benchmark evidence, Release Guard, and rollback.
5. Optimize: artifact prewarming, capacity history, Inference Lab, and advisory recommendations.
6. Expand by adapters: runtimes, clouds, external APIs, sandbox access, and training lineage stay behind narrow contracts.

Sandbox execution and training orchestration are not part of the current release. InferCrane now tracks external sandbox identity plus endpoint-scoped access and verifies signed training artifact handoffs while delegating isolation and training execution to qualified systems. Managed compute remains demand-gated because it adds billing, abuse, capacity, and support obligations—not just another adapter.

​

Explicit non-goals

InferCrane will not become another inference engine, GPU scheduler, distributed KV cache, generic router, cloud-provisioning framework, workflow engine, agent framework, or sandbox isolation runtime. Those systems remain replaceable integrations. InferCrane owns lifecycle, policy, evidence, and decision semantics.

​

How priorities are chosen

A roadmap item moves forward when it does at least one of the following:

1. Reduces time from model or existing endpoint to dependable production service.
2. Prevents an outage, leaked resource, unsafe rollout, privacy violation, or unbounded spend.
3. Produces trustworthy evidence for a decision operators currently make by guesswork.
4. Proves provider/runtime modularity without weakening the core domain model.
5. Creates repeat daily value rather than a deployment command used once.

Public milestones are intentionally outcome-oriented. Engineering checkpoint history, ADRs, and release qualification procedures remain in the repository for contributors without occupying the primary user documentation navigation.