Demo · 2:24
The finished claims system
A tour of the end product: five agents, a dashboard, an audit trail and 72 evaluations.
- The system runs five agents: intake, evidence, policy, risk and payout.
- The dashboard shows the recommendation, timeline and audit trail with confidence and warnings.
- 72 evals caught a risk agent that silently disabled a late-reporting check.
| Agent | Job |
|---|---|
| Intake | Load and normalize the claim |
| Evidence | Match images, story and estimate |
| Policy | Check coverage |
| Risk and payout | Score risk, compute payout |
Do this in your own tools, step by step. Ticks are saved on this device.
Which five agents does the system use?
Intake, evidence, policy, risk and payout.
What did the evals catch?
A risk check that had been silently disabled.
0:00 To reflect real-world complexity, I use repair damage images, repair estimates, and policies. Then build five agents: intake, evidence, policy, risk, and payout. Also, test the whole system with evaluations and package everything inside a Gradio dashboard. Then I turned the whole build into a reusable LangGraph skill and LangGraph plugin with hooks. I have my PhD with 20+ years in AI, billing, and teaching production agents across 90+ countries. Let's build together. Just before building, let me demo the final products to give you an impression.
0:30 So let me show you the final product. It has a few tabs. The claim demo, it has this damage image. There is a customer statement, the vision damage observation, and then we have a repair state and a policy. So if we say we want to run the agent team, you can see that we have a final recommendation, pay, because the image and the claim matches and the payout is this amount and the coverage ratio is 80% as deductible.
0:50 So it's just because it doesn't have any risk, it's approved. Agent timeline, the intake has been done, damage, policy, risk, and then payouts, which is says repair estimates, the coverage, 80%, the coverage cap and the cross payout. Net after deductible. The audit trail, we can see what has been done exactly, which steps and which decision has been made and exactly which evidence. And then we have the confidence, that's really important, what kind of warning if there is an image observation on certainty, we will see that.
1:40 And with the evaluations, 72 different evaluations on 15 on routing, 10 policy, 18 risk, 14 payouts and 15 on the image claim match. What did we catch by evaluations is that agent risk agent, it has a slicing silently disabled late reporting. So it's basically is just showing is this checklist. And you can imagine you can always come here and read it. Really quick, if you want a deep dive for real, what AI agents, completely for free, 30 minutes plus agents best practices and also production AI agent.
2:10 The first link is in the description. 46,000 already joined. Hope to see you there. Okay, let's build what I have just shown you. This is our build route. We go from rules, specs, data, parks, state, agents, graph, evaluations, dashboard, and finally to our skill plugins.
Get the other 11 chapters, the exam and your certificate for LinkedIn, plus all 700+ AI courses.