Zimax Networks LC · Phoenix, Arizona

Private AI inference for companies whose prompts are trade secrets.

Zimax is an inferencing company. We run managed, tiered AI inference on Microsoft Azure — with a truth-value retrieval layer, policy-enforced tenant isolation, and Microsoft's contractual commitment that your prompts are never used to train foundation models.

70/70
integration tests passing, both regions
2
regions live today (US · Mexico)
21s
measured Postgres restore RTO, 36/36 tables
0
hardcoded model names in application code

Every number on this site is backed by a versioned artifact — a test run, a drill record, or a policy file. Ask us for the receipts.

Five inference tiers. Zero model lock-in.

Applications call capability tiers — never model names. Backing models are swapped by configuration, gated by a CI evaluation contract, with tier-degrade fallback and no silent cross-vendor rerouting.

TierBuilt for
fastHigh-throughput synthesis, routing, latency-sensitive chat
standardPrimary reasoning, agent workflows, structured extraction
frontierDeep reasoning and long-context synthesis
utilityEmbeddings and reranking (1024-dim)
imageDiagram and image generation

Inference is a commodity. Knowing what to trust is not.

Every document in a Zimax corpus carries a truth-value class — Authoritative, Working, or Observed — with supersession detection and time-decay-aware retrieval. Stale documents lose ranking automatically. Superseded ones are demoted, never silently deleted.

Classified at upload

Machine classification that can never self-assign the highest class — Authoritative status requires a human, a written reason, and an append-only audit record.

Supersession detection

When a newer version of a document arrives, the old one is demoted on a decay half-life — and the system refuses to override human-asserted authority.

Proven in tests

Our executable acceptance test: a superseded document scoring 0.95 raw relevance loses to a current authoritative document at 0.70 — and remains retrievable. Demotion, not deletion.

Your prompts train no one.

Zimax inference runs on Azure AI Foundry under Microsoft's documented commitment: customer prompts, completions, and embeddings are not used to train foundation models by default — and never without your explicit instruction. Your competitive alpha stays yours.

Tenant isolation is enforced by policy at every request — not by convention. A commercial caller cannot unlock a restricted document. Every allow and deny is logged.

Read our security posture

An investor told us: “prove it.”

So we publish the receipts.

Restore drills with measured recovery times. Immutable backups verified by attempted deletion. Integration suites run against both live regions. Digest-pinned deployments where a git log answers “what is running.” We hold ourselves to the standard our truth-value system holds documents to.

The evidence page