CLASSIFICATION-AWARE BY DEFAULT

Sensitive stays home.
Heavy lifting commutes.

Most companies don't have one kind of data, so why would they have one kind of AI? The machine splits the work at the classification line: restricted inference runs on your iron, bulk training bursts to rented GPUs, and a border guard you can audit decides what crosses. Nothing sensitive commutes. Ever.

Policy-routed, not vibes-routedOne-way transfer patterns (data diode friendly)Weights travel, data doesn'tEvery crossing logged
Data diodes & one-way transfer Cross-domain patterns Signed artifact promotion Split RAG (local index / cloud reasoning on redacted context) Policy-as-code (OPA) Full crossing audit log Data diodes & one-way transfer Cross-domain patterns Signed artifact promotion Split RAG (local index / cloud reasoning on redacted context) Policy-as-code (OPA) Full crossing audit log

All-or-nothing is a procurement position, not an architecture.

Go all-cloud and your export-controlled engineering data is grounded, so the most valuable half of your AI program never takes off. Go all-on-prem and you're buying H100s to summarize marketing emails. The companies getting this right stopped asking “cloud or on-prem?” and started asking “which data, which direction, under whose key?” That's a routing problem. The machine is very good at routing problems.

2 systems ≠ 2x cost
shared eval, shared registry, one pane
~80/20
most token volume is unrestricted; most VALUE is restricted
0
manual decisions at the boundary, policy decides, humans audit

Four proven traffic patterns. Your mix will use two or three.

Nobody runs all four on day one. We start with the pattern your data mix actually needs and add the rest as your program grows.

Train there, infer here.

Deep domain fine-tune needed, but production queries must never leave your fence line.

Scrubbed / synthetic corpusthen
Cloud GPUs, heavy fine-tunethen
Signed weights courier homethen
Inference runs on-prem, on real data

Crosses: Sanitized training data out (once), weights in.

Never: Production queries.

Which pattern fits us?

Route by classification.

Every request enters a policy gate, this page's hero, in production.

Request enters the gatethen
Unrestricted → frontier cloud APIthen
Restricted → local fine-tuned model

Crosses: Only requests the policy marks cloud-eligible.

Never: Restricted requests, the router reads the data's markings, not the user's mood.

Which pattern fits us?

Split RAG.

One knowledge base, two audiences, some questions cleared for a frontier brain, some not.

Vector index + documents stay localthen
Unrestricted query → cloud model reasons over redacted contextthen
Restricted query → local model, same index

Crosses: Retrieved-and-redacted context, for unrestricted queries only.

Never: Raw documents or the index itself.

Which pattern fits us?

Burst with a diode.

Shops with real cross-domain hardware, we design to your CDS, not a whiteboard fantasy.

Nightly batch ships one-way (diode-friendly)then
Cloud compute runs the analyticsthen
Results return through a verified gate next cycle

Crosses: Batch data out (one-way); verified results back.

Never: Anything bidirectional, real-time, or unverified.

Which pattern fits us?

The Classification Router

You become the border guard. Then you learn why you shouldn't be.

1 · Inspect mode

Click any queued request to see its grade, its content, and the policy's verdict.

Pick a request above to see how the gate reads it.

Want the full grade walk instead?

What are you holding?

Two kinds of data, two kinds of iron, one policy.

Your mix has both restricted and unrestricted data. A classification-aware gate routes each request to the right side automatically.

Illustrative. Your compliance officer outranks this widget, bring them to the assessment.

0
policy violations across every deployment
11d → 61h
70B fine-tune: on-prem est. vs. cloud burst
100%
of crossings logged & attributable
1
pane of glass for both sides

“Edge telemetry stays on the floor where it's generated; the heavier fleet-wide analysis bursts to the cloud overnight and comes back signed. We finally stopped debating where the AI should live, the gate decides, and it's never been wrong.”

Data-center operator: ThermaRay edge + cloud telemetry deployment

Frequently asked

Draw the line once. Enforce it forever.

A 30-minute deployment assessment: your data mix, your traffic patterns, and a routing policy the machine can actually enforce, not just diagram. Run by people who've shipped it, not sold it.

30 min with a founderCustom ROI estimateNo commitment

Typically responds within 4 hours

You own every line, including the routing policy.