A decision framework that puts approved data movement ahead of a provider preference.
Local is a complete architecture decision
For a fully local proposal, we would enumerate every model, extraction tool, identity dependency and operational service. The bank should know which components need updates or external access. Hardware selection follows the validated workload and model choices.
Cloud infrastructure and cloud models are different choices
A cloud-hosted application can use a dedicated model server or a provider API. Those arrangements have different contracts, processing paths and operating responsibilities. Our comparison would make the application, data and inference locations explicit.
Hybrid needs a clear traffic map
We would specify which requests cross each boundary and which are refused. Include prompts, attachments, results, logs and support access. A hybrid label should not hide a data-transfer decision that the bank has not approved.
Size and price a real workload
Our sizing worksheet would record model, precision, context length, concurrent requests, latency target and document volume. Benchmark the candidate configuration, itemize recurring costs and document what a higher demand level would require.
Make the next step specific.
Start with a bounded assessment, a named business owner and a clear decision to be made. The final scope and responsibilities belong in the engagement agreement.
Explore the related service