Inference pilot · scoped and quoted before work begins
Prove one inference workload before you commit to capacity.
For teams considering Australian AI infrastructure. Vision and video come first; LLM pilots use their own criteria.
- 1 · AgreeScope and controls
A scope workshop fixes one application, an authorised dataset, the test environment, success thresholds, responsibilities, price and timeline.
- 2 · MeasureReproducible evaluation
A documented baseline, then the agreed tests. Vision and video: detection quality, event latency, data transferred and cost per camera-hour. LLM: fix the model, precision, context, output length and arrival pattern first; then measure task quality, time to first token and cost per accepted task.
- 3 · DecideA written decision
Results with limitations, an acceptance meeting and a recommendation to accept, revise or stop. Next steps are scoped separately: a customer-owned deployment or, once sites are accepted, reserved InferaStack AI Grid capacity.
Not included unless agreed in writing: production SLA, ongoing hosting, hardware purchase, model training or fine-tuning, unrestricted sensitive-data processing and application development.
The proposed validation environment
| Stage | What it involves | Status |
|---|---|---|
| Lab | An evaluation environment selected for the agreed workload, authorised data and success criteria. | Proposed. Lab results show software behaviour only, not site GPU performance. |
| Service readiness | Assess operational requirements, data handling and service continuity before any production commitment. | Scoped for each pilot; subject to validation. |
InferaStack AI Grid is a proof of concept in technical validation. Pilot results inform its validation plan →
Agree data handling before testing
We agree where data is processed, who can access it, what is kept and when it is deleted. Running the service from Australia does not by itself mean data is processed only in Australia. Sensitive data needs an environment approved for it. The InferaStack Gateway is in development.
Bring one workload and its success criteria.
Tell us the application, data location needs, expected load, budget and timing.