VeriVeri API – Fact verification for AI output

Every AI output. Verified. VeriVeri API is a fact checker: it verifies text — written by an AI or by a person — against the documents it was supposed to be based on, detects hallucinations and factual errors, and tells you what is supported, what is wrong and what a person should look at. Human review time drops dramatically, and your AI systems stay under control. It deploys as a managed application into your own Azure subscription: your data, your region, your models.

Where VeriVeri fits

• Customer-facing chatbots and digital assistants — every answer checked before it reaches the customer.
• AI analysis, summaries and reports where the cost of an error is high: internal control and audit, finance and investor relations, external communications, regulatory filings, management reporting, business planning and strategy.
• Long-running agentic systems — grounded decisions at every step, less quality drift in complex, data-heavy workflows.
• Document Q&A over contracts, policies and procedures.
• Human-written material too: press releases, board papers and documentation checked against their sources.
• Evaluation pipelines and release gates for AI features — an auditable quality signal.

How it works

• Send the grounding context and the text to check to POST /v2/verify.
• Several models from your Azure AI Foundry — GPT, Claude, DeepSeek, Grok, Kimi, your choice — judge every claim independently and in parallel; long documents are chunked automatically.
• A consensus rule you control turns their votes into one verdict — correct, wrong or needs_review — with each flagged passage quoted, an explanation, a proposed fix, confidence and model cost.

What is included

• REST API with OpenAPI reference, idempotent requests, per-key limits, budgets and expiry.
• Operator dashboard: Playground, Audit log with export, Spend by model, Backends and Models, versioned verification prompts per domain (general, finance), Privacy & Security.
• MCP server with Microsoft Entra sign-in for agents and assistants (verify_text, estimate_cost).
• Failover across Foundry workspaces and optional external providers — OpenAI, Anthropic, any OpenAI-compatible provider or a LiteLLM AI gateway — with 429 back-pressure handling and a degraded flag when a partial pool decided.
• Privacy controls: payload storage on or off, PII masking at rest, retention purge, right-to-erasure endpoint, append-only operator trail.
• Onboarding and optimisation by our customer success team (opt in during deployment): model setup, pool tuning for your use cases and scenarios, and LLM capacity planning — quotas, concurrency and response times — so peak loads hold.

Runs in your subscription

Container Apps, PostgreSQL and Key Vault behind private endpoints, Log Analytics and Application Insights with alerts — created by the wizard in the region you choose. EU, US or Global: you decide where the workload and the models run, and support email and probes stay there.

Nothing leaves your tenant unless you add an external provider or send a support case; Northloop's access is limited to the managed resource group, for support and updates only.

Included in the monthly fee, on when you need them: private-only API exposure, IP allowlist, Microsoft Entra SSO for operators, and geo-redundant backups on the production database tier.

What you need

• An Azure subscription and models in your own Azure AI Foundry — or deploy first and connect OpenAI, Anthropic, any OpenAI-compatible provider or a LiteLLM AI gateway afterwards.
• Tested model configuration (pool optimisation last run August 2026):
  US or Global — short inputs: gpt-5.4-mini, DeepSeek-V4-Flash, gpt-5.6-luna, grok-4-20, claude-sonnet-5 · long inputs: gpt-5.4, gpt-5.6-terra, gpt-5.6-sol, claude-opus-5.
  EU Data Zone (Data Zone deployments, EU region) — short inputs: gpt-5.4-mini, DeepSeek-V4-Flash, gpt-5.6-luna · long inputs: gpt-5.4, gpt-5.6-terra, gpt-5.6-sol.
  Minimum two models from two providers per pool — the wizard checks and tells you.
• A deployment from scratch takes about 30 minutes and you are ready for your first verification — paste one Foundry key, pick a privacy preset.
• Azure bills the infrastructure (a minimal install idles at a few tens of euros a month); your Foundry bills model usage.

Plan

A monthly subscription with no commitment: a flat fee per deployment on your Azure invoice, with no limit on verifications.

Good to know

Verdicts are probabilistic aids for human review, not determinations of truth. Support from the dashboard's Support page or support@veriveri.io, acknowledged within one business day; policy at https://veriveri.io/legal/support.