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

Cautelia.

Takes the report, checks the policy, hands it to an AI adjuster. A portfolio project on how real agents are built.

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

Cautelia takes the report by voice or text, finds the answer in the insurer’s policy and leaves the decision to an AI adjuster that explains it. A portfolio project.

Try the demo

Why this exists

Not another wrapper.

Calling a model is the easy part. Cautelia shows what most demos skip: what happens when the provider fails, where the state lives, where each figure comes from, and proof that it works.

  • Every citation verified
  • Insurers kept apart, tested
  • Data redacted in logs and traces
  • An AI adjuster decides coverage
One case, step by stepScripted example, not a recording
  1. Takes the report4 turns
  2. Reads the policy2 clauses
  3. Checks the citationverified
  4. Reviews the case3 checks
  5. An AI adjuster decideswith the trace

How it works

One orchestrator. Specialist agents. Small shared state.

Code, not the model, decides which node handles each turn. Nodes return structured results, never raw reasoning, so the context stays small and the trace reads top to bottom.

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

    Orchestrator

    Routes each turn in code and decides when the AI adjuster steps in.

  • 02

    Intake

    Gathers who, what, when and where while the caller talks, in any order.

  • 03

    Policy verification

    Cites only clauses that exist, belong to that policy and are in force. “Nothing found” and “could not check” are different answers.

  • 04

    Decision

    Applies the insurer’s versioned, tested rules. The model never picks the category.

  • 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
  • The hard part is knowing when someone has finished. Cautelia waits for meaning, not silence: whoever dictates a plate pauses mid-number.
  • The caller can interrupt. The agent stops mid-sentence, keeps the state, and picks up where the caller left it.
  • The heavy work (policy checks, rules, compliance) never runs inside the turn, so answer time does not depend on how many rules an insurer has.

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?

recordlocation✓time✓injuriesnone ✓other partypending

Guardrails

The agents gather. An AI adjuster decides.

A report is intake and coordination, not judgment. Cautelia never says something is covered, never denies it and never mentions amounts: it hands the case to an AI adjuster that decides and explains each step with its sources.

  • Before the model: outside text enters as data, never inside the prompt.
  • In flight: every tool call is validated against its schema.
  • Before anything leaves: every message, errors included, passes one exit guard.
  • Above all: if someone reports a serious injury, the agent stops collecting data and gives the emergency phrase.

Evals

Every change is tested. What fails is shown.

Scripted conversations with known answers run on every change. Pass or fail per case, Spanish and English apart, never averaged. A failing case stays visible: it is never removed to make a number go up.

Every call leaves a nested trace, so a claim can be followed from the first word to the closed record.

What gets measured

  • ✓Latency per turn, end to end01
  • ✓Interruptions handled without losing state02
  • ✓Completeness of the FNOL record03
  • ✓Latency per turn, and stop time on interruption04
  • ✓Zero personal data in logs or traces05

See it run.

Describe a made-up crash, by chat or by voice, and see what it does. Nothing to install.