How it works

One layer.
Between your apps and the AI.

Your apps don't change. The AI vendors don't change. Vorana sits in the middle — checking, scoring, and recording every answer that flows through.

Four steps

Connect, configure, govern, improve.

1 Connect

Point your apps. Change one line.

Vorana speaks OpenAI's wire format, so most teams flip base_url and a model string and they're done. Use the typed .NET SDK if you'd rather stay in-process.

  • OpenAI-compatible drop-in
  • SDKs for .NET, Python, TypeScript, JVM
  • Streaming, tool-calls, and JSON-mode all preserved
app.py · diff vs. main
 from openai import OpenAI
 
 client = OpenAI(
-    base_url = "https://api.openai.com/v1",
+    base_url = "https://gateway.vorana.ai/v1",
+    api_key  = os.environ["VORANA_KEY"],
 )
 
 resp = client.chat.completions.create(
-    model = "gpt-4o",
+    model = "vorana:pipeline.faq_assistant.v1",
     messages = [...],
 )
2 lines changed no other code touched, streaming preserved
2 Configure

Pick your guardrails. In plain language.

Open the admin console, choose vendors per tenant, lock regions, set monthly budgets, turn on PII and PHI redaction. Saved settings apply on the next request — no redeploy.

  • Per-tenant vendor allowlists
  • Region locks & data-residency rules
  • Hard budget caps with alert thresholds
vorana / admin · tenant_acme
Vendors allowed
Azure OpenAI OpenAI Anthropic Bedrock
Region lock
eu-west-1 no egress outside EU
Monthly budget
$10,000 / tenant · alert at 80%
Required redaction
PII PHI PCI Custom
Saved · applies on next call
3 Govern

Every call. Every check. Automatically.

Each request flows through the pipeline you configured: validate inputs, retrieve grounding, run consensus across providers, score, gate by policy, then sign and append to the audit log — before the answer reaches your user.

  • Validation, retrieval, consensus, score, policy, audit
  • Failures route to fallbacks, not errors
  • Decisions captured at every step
vorana / live · r_a31f…
input.validateschema ok
retrieval.hybrid3 sources cited
llm.consensusazure_openai & anthropic · 0.92
scoring.compositejudge 0.91 · cache miss
policy.gatepass · eu-west region ok
audit.appendCMK encrypted · signed
200 OK latency 1.4s · cost $0.0021 · run_id r_a31f…
4 Improve

One dashboard. Quality, cost, risk.

Watch quality scores rise as cache warms. Watch per-team cost fall as cascades route to cheaper models. Replay any low-confidence run against today's pipeline to verify a fix — without ever touching production traffic.

  • Quality and cost trends per pipeline / team
  • Flagged runs with one-click replay
  • Export evidence for SOC 2, HIPAA, GDPR
vorana / dashboard · last 7 days
Quality
0.93
↑ 4% wk/wk
Cost / call
$0.018
↓ 32% wk/wk
Audit coverage
100%
CMK signed
3 runs flagged for review
r_2c9d… low_conf 0.74 support-copilot Replay
r_71b2… disagree claims-intake Replay
r_4c8a… cost_spike advisor-agent Replay
Inside the layer

Six things every call gets — for free.

Validation

Wrong inputs get rejected before they reach the AI. Wrong outputs get caught before they reach the user.

Grounding

Pulls in your documents, FAQs, or knowledge base — so answers come from your truth, not the model's memory.

Cross-checking

Asks two or three AIs the same question. Only ships the answer when they agree — or sends a human-review signal.

Policy

Region-locked routing, vendor allowlists, PII masking. Your security team writes the rules — Vorana enforces them.

Caching

Same question? Same answer — for free. Real workloads see cost cut by half once cache warms up.

Audit

Every call captured — inputs, outputs, decisions, citations. Replayable, exportable, immutable.

Where it runs

Your cloud. Our cloud. Or air-gapped.

Managed cloud

We run it. Multi-region, 99.9% SLA, hosted dashboard. Easiest to start.

Your cloud

Self-hosted on Azure, AWS, GCP, or on-prem. Full data sovereignty. Same product.

Air-gapped

Disconnected, signed-bundle distribution for high-trust environments. No phone home.

Want a walkthrough?

30 minutes — we'll show you the dashboard, the policy editor, and a real call flowing through.