Grokking the Forward Deployed Engineer Interview
Vote

0% completed

​

The Smallest Version That Delivers Value

What the Smallest Version Is

Choosing the Queues

What Was Cut, and Why

What Could Not Be Cut

Thin Slice or Layer by Layer

Common Mistakes

Using This in an Interview

Key Takeaways

Practice Questions

Title image: of the six queues, glass and minor damage are chosen for the first version, about half of all claims.
Title image: of the six queues, glass and minor damage are chosen for the first version, about half of all claims.

The first version of the insurer's routing system did five things:

  • It read every new claim through the claims API.
  • It suggested a queue and a confidence score for every claim.
  • It routed glass and minor damage claims automatically when the score was above the threshold.
  • It wrote every decision to the decision log.
  • It could turn automatic routing off with one setting.

It left out automatic routing for the other four queues, a new adjuster screen, damage assessment, and payments. That list matters as much as what was built.

What the Smallest Version Is

The smallest version that delivers value is the least the team can build that still changes a user's work and produces a measured result.

A thin slice is narrow but complete: it runs every step from input to result, but only for a limited part of the work. The insurer's slice ran every step for every claim, but routed only two queues automatically.

Choosing the Queues

The FDE compared the six queues on two numbers: the share of claims each receives, from the queue history, and the cost of a wrong automatic decision, from the adjusters.

QueueShare of claimsClaims a dayCost of a wrong automatic decision
Glass20 percent240Low: about 1 day lost while an adjuster reroutes it
Minor damage30 percent360Low: about 1 day lost
Major damage20 percent240Medium: a wrong repair estimate may be started
Injury12 percent144High: a legal deadline for contact may be missed
Theft10 percent120Medium: the police report check and the theft review start late
Possible fraud8 percent96High: a fraudulent claim may be paid
The six queues by share of claims, with glass and minor damage chosen for the first version because they receive half the claims at the lowest cost of error.
The six queues by share of claims, with glass and minor damage chosen for the first version because they receive half the claims at the lowest cost of error.

Glass and minor damage together receive 50 percent of claims, about 600 a day, and a wrong decision in either costs about 1 day. No other pair offers as much volume at as little risk, and photos decide most of their claims.

The value can be estimated before building. If 70 percent of the 600 pilot claims are routed automatically, 420 claims at 5 minutes each is 2,100 minutes, or 35 adjuster-hours a day.

What Was Cut, and Why

Each cut passed the same test: the version left still changes the adjusters' work, meets the guardrails, and can be measured.

  • Automatic routing for the other four queues. The system still suggested a queue for every claim, so the team could measure the suggestions.
  • A new screen for adjusters. The suggestion and its score went into a notes field that the claims system already showed. A new screen would have taken about a week and changed nothing about the decision.
  • Damage assessment and payment. These are separate decisions with their own data and higher risk.
  • Changes to the model itself. The first version used the model as the insurer bought it. The FDE improved the inputs instead: the photos, the converted reports, and the facts list.

What Could Not Be Cut

The decision log was kept because the regulator requires it. Without it, the first version could not legally route a real claim.

A switch to turn automatic routing off was kept because the pilot needed a safe way to stop. If agreement had fallen, one setting would have sent every claim back to the adjusters within minutes.

The rule is simple: cut any part that the first version can work without, unless the law requires it or the team needs it to stop safely.

Thin Slice or Layer by Layer

The other way to build is one layer at a time: extraction for all six queues first, then the rules, then the routing. Nothing would reach an adjuster until every layer was done.

Two ways to build: a thin slice runs every step for two queues in week 4, while building layer by layer reaches a user only after every layer is finished.
Two ways to build: a thin slice runs every step for two queues in week 4, while building layer by layer reaches a user only after every layer is finished.

The thin slice reaches real users sooner, and real use shows problems that no test set can. In the pilot, adjusters noticed that night photos often scored low, and the team fixed it. Building layer by layer would have found the problem weeks later.

Common Mistakes

The first mistake is a version that is too wide. Routing all six queues at once would have put injury and fraud claims at risk when the team knew least about the system.

The second is a version that is too thin. Suggestions alone, with no automatic routing, would have saved no hours. Adjusters would have used them as a hint, and the outcome could not be measured.

The third is cutting what cannot be cut. A pilot without the decision log would have had to stop when the regulator asked for its records.

Using This in an Interview

When asked what you would build first, name a thin slice, show its value with arithmetic, and give a reason for each cut.

"I would start with glass and minor damage, about 600 claims a day and the cheapest to get wrong. At 70 percent automatic, that saves about 35 adjuster-hours a day. I would keep the decision log and an off switch, and leave the other four queues as suggestions."

Expect two follow-up questions:

  • "Why not include injury, since those claims are urgent?" Urgency raises the cost of a wrong decision, so injury waits until the pilot has shown the real error rate on safer queues.
  • "What would you build second?" Pick the next queue by share and cost of error, and name the pilot measure that must stay on target first.

Key Takeaways

  • The smallest version that delivers value changes a user's work and produces a measured result.
  • Choose a thin slice by volume and by the cost of a wrong decision.
  • Cut any part the first version can work without, unless the law requires it or the team needs it to stop safely.
  • A thin slice reaches real users sooner, and real use finds problems that a test set cannot.

The next lesson, Writing a One Page Scope, puts the slice, the measures, and the cuts into one document that the customer signs.

Practice Questions

1. Why did the first version suggest a queue for every claim, even for queues it did not route automatically?

<details><summary>Show answer</summary>

Because the suggestions could be measured without any risk. Adjusters still decided those four queues, and comparing suggestions with their choices showed how the system would perform before automation.

</details>

2. Why was the decision log kept in the first version when the new adjuster screen was cut?

<details><summary>Show answer</summary>

Because the regulator requires the log, while the screen only added convenience. Without the screen, adjusters still saw the suggestion in an existing notes field.

</details>

3. A teammate proposes perfecting extraction for all six queues before anything goes live. What is the main risk?

<details><summary>Show answer</summary>

No user sees the system until every layer is finished, so problems from real use appear late. The night photo problem would have been found weeks later.

</details>

Reading Progress

0%


Vote for new content

On This Page

What the Smallest Version Is

Choosing the Queues

What Was Cut, and Why

What Could Not Be Cut

Thin Slice or Layer by Layer

Common Mistakes

Using This in an Interview

Key Takeaways

Practice Questions