ZXO / LOCAL AI PILOT
On-premises AI pilot
for U.S. teams.
Test private document Q&A or another agreed workflow on a workstation or your compatible host. Your team can ask about approved manuals, procedures or code and inspect the source excerpts behind the answers.
01 / WHAT YOU GET
A configured system and a tested workflow.
Compute and runtime
A workstation selected for the pilot or setup on your compatible host. We preinstall and tune a supported driver and runtime, such as NVIDIA CUDA or AMD ROCm, and choose Ollama or vLLM based on the hardware, operating system and workload.
We assess MLX, multi-GPU inference and distributed serving separately for the chosen platform.
One model and workflow
One agreed model is preloaded and tested. Use a local web interface to ask questions about approved documents or code and review source excerpts. An optional assistant workflow can use the specific data and tools you approve.
Results and handoff
Receive the hardware and software version sheet, tested settings and results, plus startup, restart and handoff instructions. We agree on the test questions and target usage before integration.
Offline option. We stage approved models and dependencies during provisioning. When offline operation is part of the scope, we test the agreed workflow with outbound access blocked before handoff.
02 / FICTIONAL HANDBOOK EXAMPLE
Check the answer against the manual.
Question
How much rear exhaust clearance does the Falcon F27 require?
Demo answer
The Falcon F27 requires at least 80 mm of unobstructed rear exhaust clearance.
“The Falcon F27 rear exhaust requires at least 80 mm of unobstructed clearance. Measure from the grille to the nearest solid surface.”
Source excerpt from falcon-f27-cooling.txt, a fictional engineering handbook used in the ZXO demo.
When the manuals have no answer
Question
What is the Orion Q92 battery replacement interval?
Demo answer
The supplied documents do not contain the answer.
This demo passed eight scripted checks on a Windows test host with Qwen3 8B, including one question the handbooks could not answer.
How we assess your pilot
Before setup, we agree on the documents, test questions, target response time and expected simultaneous users. For document Q&A, we check that answers cite the supporting passages and identify questions the approved documents cannot answer. The handoff records which checks passed and where the workflow needs more work.
03 / MODEL AND HARDWARE EXAMPLES
Match model size to the workload.
These profiles are starting points for evaluation. We confirm hardware availability, model licensing and runtime compatibility before finalizing a configuration.
W2 planning example / 96 GB VRAM
Single NVIDIA RTX PRO 6000 Blackwell Workstation Edition. Candidates include 4-bit builds of DeepSeek R1 Distill Llama 70B or Llama 3.3 70B Instruct, and Qwen2.5 Coder 32B Instruct for code workflows.
OEM GB10 example / 128 GB shared memory
Evaluate a compatible Arm Linux stack on an OEM GB10 system. The operating system, CPU and GPU share memory. We check the selected runtime and model build, then test context length and concurrency together.
What the memory figures mean
A 70B model needs roughly 140 GB for BF16 weights alone, beyond a single 96 GB GPU. At 4 bits per weight, the theoretical baseline is about 35 GB. Quantization metadata, runtime overhead and the attention cache add to that figure. We test response time, context length and concurrent requests on the actual system.
04 / TYPICAL PILOT SCHEDULE
Three steps over four weeks.
Illustrative schedule for U.S. projects, starting after we confirm scope and the required equipment is available. The project agreement sets the actual schedule, delivery and support.
Week 1 / Assess and match
Review the workload, approved data, expected usage and network requirements. Select the host, model, runtime and acceptance tests.
Weeks 2–3 / Integrate and test
Configure the host, preload the model and runtime, and run the agreed tests for source accuracy, response time and usage. Verify offline operation when included.
Week 4 / Deliver and tune
Hand over the configured system or compatible host setup, test results and operating notes. Arrange U.S. delivery and on-site or remote tuning where included in the project scope.
05 / WHAT TO BRING
Give us enough to scope the test.
- The task, typical questions and an example of a correct answer.
- Document formats, approximate volume and any access or network constraints.
- Your preferred hardware or existing host, expected simultaneous users, acceptable response time and target start.
Describe your files in the inquiry. Share approved sample documents only after agreeing on a test environment and transfer method.
LOCAL AI / PROJECT INQUIRY
What should your team be able to find?
Tell us what you want to run, how many people will use it, and where it needs to operate. We will review the hardware, workflow, timeline and pricing for your pilot.
