Cautelia.

A multi-agent AI system for insurance quoting, underwriting and claims, with voice in real time. Built as a portfolio project to show how production agents are actually made.

Try the demo or see how it works

Someone just had a crash and does not know what their policy covers.

Cautelia is a multi-agent AI system for insurance quoting, underwriting and claims, with voice in real time. A portfolio project built to show how production agents are actually made.

Try the demo

Why this exists

Not another wrapper.

Reviewers have seen the chatbot around an API a hundred times. Cautelia was built to show the parts that project skips: failure handling, structured state, real integrations, and proof that it works.

  • Handles interruptions
  • Structured handoffs
  • Golden-set evals
  • Human on every decision

How it works

One orchestrator. Specialist agents. Small shared state.

Each turn goes to the right specialist. Agents hand back structured results, never raw reasoning, so the context stays small and the trace stays readable.

Architecture: voice layer to orchestrator, orchestrator to specialist agents, agents to tools, orchestrator to human handoff. Nodes Tools Voice layer Orchestrator Intake Policy verification Decision Compliance Policy corpus Rules engine Human handoff
  1. Voice layer
  2. Orchestrator
  3. Specialist agents
  4. Tools
  5. Human handoff
  • 01

    Orchestrator

    Routes each turn in code.

  • 02

    Intake

    Gathers the facts while the person talks.

  • 03

    Policy verification

    Only cites clauses that exist.

  • 04

    Decision

    Applies the versioned rules.

  • 05

    Compliance

    Reviews the case before it closes.

Voice

Fast enough to feel like a call. Patient enough not to interrupt.

  1. Streaming speech to textstreams
  2. End-of-turn detectionwaits for meaning
  3. Orchestratorroutes
  4. Tool callsvalidated
  5. Streaming text to speechstreams
  • End-of-turn detection is the long pole. Cautelia waits for meaning, not for silence.
  • The caller can interrupt. The agent stops mid-sentence, keeps the state, and picks up where the caller left it.
  • Every stage streams. Nothing waits for the whole answer.

Scripted example, not a recording

  1. Caller

    Hi, I had a fender bender this morning, nobody's hurt.

  2. Agent

    Glad everyone's okay. I'll open the claim while we talk. Where did it happen?

  3. Caller

    On the highway, near exit 12, around eight.

  4. Agent

    Got it: highway, exit 12, about 8 a.m. Was the other driver's information exchanged?

record: location ✓ time ✓ injuries: none ✓ other party: pending

Guardrails

The agents gather. People decide.

A first notice of loss is intake and coordination, not judgment. Coverage and money decisions belong to the carrier's rules and its adjusters. Cautelia is built on that line.

  • Before the model: input checks, personal data handling, scope limits.
  • In flight: every tool call is validated against a schema. No free-form writes.
  • After the model: anything that touches coverage or money goes to a person, with the trace attached.
  • And above all of it: the emergency phrase.

Evals

A golden set, a score, and an honest list of failures.

Every change runs against scripted calls with known right answers. The report is a pass or a fail per case, and the cases that broke stay in the open.

Every call leaves a trace: each model call, each tool call, each handoff, nested so one claim can be followed from the first word to the final record.

What gets measured

  • ✓Latency per turn, end to end01
  • ✓Interruptions handled without losing state02
  • ✓Completeness of the FNOL record03
  • ✓Correct handoffs to a person04

Try it

Press and hold to verify.

Press and hold.

  1. Call received
  2. Facts captured
  3. Quote drafted
  4. Referred to a person
  5. Record complete

That's the whole product.

FAQ

The questions reviewers actually ask.

Is this just a wrapper around a model?

No. The model is one component. The work is in the orchestration, the structured state, the tools, the voice pipeline, the guardrails and the evals. All of it is on the table in a walkthrough.

Does it work when people interrupt or talk over it?

That is the point of the voice layer. The caller can interrupt, the agent stops, keeps the state and picks up where the caller left it. It is also the hardest part, and the eval set tracks it.

Who decides coverage?

People do. The agents gather facts, draft and refer. Anything that touches coverage or money goes to a person with the trace attached.

Is this a real insurer's product?

No. Cautelia is a demonstration project built for a portfolio. The insurance flows are real patterns. The carrier, the data and the imagery are not.

Can I see the code or a live run?

Yes. Message me on LinkedIn and I will walk you through a live call, the traces behind it and the code.

See it run.

A twenty-minute walkthrough: a live call, the traces behind it, and the code.

Try the demo or message me on LinkedIn