A small, opinionated piece of software that sits between your firm's documents and the person on the phone. Three outputs. Every answer cited. Nothing leaves the box.
Verdict is deliberately simple. The whole stack fits on a single 32GB box. The whole stack also fits in your head, which is why your compliance team will sign off on it.
Dense embeddings for semantic match. BM25 for terminology fidelity (your firm uses "trailing commission" — Verdict knows that's not a typo). Re-ranked by cross-encoder. Every passage carries its source pointer through the entire pipeline.
We run a fine-tuned Gemma 7B on-device. No GPT-4. No Claude. No cloud inference of any kind. Weights are versioned and signed; you can pin a specific build for an entire fiscal year if your auditor asks you to.
Handle / Transfer / Escalate. The model is constrained to a JSON schema — no free-form prose, no "as a language model." Every field is required. Every citation is a real passage in your corpus. Hallucination is structurally suppressed, not hopefully prompted.
The whole point is that your IT team can support this without learning anything new. We wrote it to look like every other line-of-business app on your network.
One Windows or Linux box. 32GB RAM minimum, 64GB recommended. CPU-only mode supported for firms under 50 seats. Comes as an MSI or a signed Docker bundle.
Ship a sealed appliance for firms that physically can't connect to the public internet. Updates arrive via signed USB. We've done this for one customer already; it's not theoretical.
SAML, OIDC, Azure AD, Okta. Group-level corpus permissions: a junior rep sees the rep manual, the compliance officer sees everything. SCIM provisioning included on Standard and up.
There's a list of things people ask us for that we won't build. We think it's more honest to publish the list.
Verdict answers structured questions tied to a customer interaction. It will not write your performance reviews or summarize your inbox.
If the local model can't answer with high confidence, Verdict says so. It does not silently call out to GPT-4. There is no toggle for this.
The compliance team writes the rules; Verdict reads them faster than a human can. Those are very different jobs.
LoRA fine-tunes on the Custom tier are scoped to your tenant, encrypted, and never aggregated across customers. Even your aggregate metrics don't leave the box.
Send us two redacted policies. We'll be back with a working demo in 48 hours.