Crudo · On-device AI support
Keep your agents, and your data, in-house.
Crudo is the AI customer service platform built on local compute. Custom agents built for your knowledge and systems, to provide trustworthy, reliable support.
What Crudo is
Four things, held at the same time.
On your hardware.
Crudo runs open-weight models on your own machines. Customer data never leaves your infrastructure, and there's no metered inference bill.
Grounded, not guessing.
Every factual claim is checked against your knowledge base, and every response is vetted by an independent model to ensure factual accuracy. When it can't back a claim, it refuses instead of inventing one.
Real work, not just chat.
Kanpachi looks up accounts, processes eligible actions, files tickets, and updates cases in the systems your team already uses.
Stays in the loop.
When a conversation needs a human, Kanpachi opens the ticket, briefs the person, and keeps the case moving, instead of disappearing.
How it works
One agent, from first message to resolution.
Base model, knowledge, and action, composing into a single agent.
- 01
Understand the request.
Every message is triaged first: what does the customer want, and should Crudo handle it? Off-topic and manipulation are caught here, before a cycle is spent.
- 02
Ground the answer.
Kanpachi retrieves the most relevant material from your knowledge base and reasons over it, using your documented policies, not a model's memory.
- 03
Take real action.
Look up an account, process an eligible refund, check a known issue, file a case, all through secure connections. Irreversible actions require explicit confirmation.
- 04
Validate before sending.
Every reply is checked against the source material actually retrieved. Unsupported claims are stripped or corrected. If it can't be grounded, it revises or hands off rather than guess.
- 05
Escalate, and stay in the loop.
If a human is needed, Kanpachi opens a ticket with a written brief, tells the customer what's next, and keeps the case moving. Human decisions are relayed back, validated the same way every reply is.
Runs on your hardware
Your models. Your machines. Your data.
Crudo runs open-weight models on infrastructure you control. There's no external inference API in the loop. This isn't a deployment option, it's the foundation.
Private by architecture
Every message, record, and transcript is processed locally. Nothing is sent to a third-party model provider or used to train an outside system. The most reliable way to keep data private is to never transmit it.
- No customer PII crosses your network boundary.
- Nothing retained or used to train external models.
- Data-residency and compliance satisfied by default.
Cost-stable by design
Cloud AI bills you per token, per message, per resolution, costs that scale with the very thing you want to grow. Crudo runs on hardware you provision once. Volume goes up; your inference cost doesn't.
- No per-conversation or per-token fees.
- Costs are capacity, not a tax on success.
- Right-size the hardware to your load.
The core principle
Local compute gives you command over your entire support ecosystem. The trust and security of on-device models is exactly what lets your agents do more, safely handling real work that a cloud agent never could.
Grounded knowledge
Answers you can trace to a source.
Before any response reaches a customer, Crudo compares the agent's claims against the material it actually retrieved. Supported claims pass. Unsupported ones are removed, or the response is revised. If it can't be grounded, the agent hands off rather than send a guess.
Finds the right thing
The exact passage that answers a question, not a vague neighborhood of it.
Checks its own work
Automatic on every message. When grounding fails, it escalates instead of improvising.
Your team manages it
Propose and edit through a guided flow. Every change validated and health-checked before it ships.
Integrations
A conversation becomes resolved, tracked work.
CRM & case management
- Opens and maintains cases automatically.
- Checks history so recurring issues are recognized.
- Classifies into your own taxonomy for clean reporting.
- Cross-links escalations and their engineering tickets.
Known issues & engineering
- Checks your engineering tracker first.
- Gives customers the real status and workaround.
- Attaches customer reports to existing issues.
- Files structured tickets, pre-written with context.
Built for your stack
Beyond the systems above, we build custom integrations to fit the tech stack you already run. You'd be surprised how much Crudo's agent can do once it's wired into the tools your team lives in.
Human-in-the-loop
Escalation isn't the end of the agent's job.
A handoff is one more step the agent manages. Kanpachi briefs the human, tells the customer what's next, relays the validated decision back, and keeps every reply flowing into the same case.
Security is the foundation, not a feature.
The architecture assumes sensitive data, real customer actions, and the reality that an agent will sometimes be uncertain, and it's designed to fail closed. The safe path is the default path.
- 01No external inferenceData doesn't leave your infrastructure.
- 02Confirmation-gatedIrreversible steps require an explicit yes.
- 03Grounded or nothingUnsupported claims don't get sent.
- 04Human authority stays humanThe agent can't grant itself permissions.
- 05Auditable by designActions and reasoning are recorded.
See the whole loop, running locally.
A short demo walks a real conversation end to end: a grounded answer, a real action in a connected system, and a clean escalation where the agent stays in the loop. No customer data required.