THE FDE LEARNING PATH
Build the system.
Own the outcome.
A claims team is doing the same work twice. Learn to find the cause, build a reliable workflow, and defend your decisions in an FDE interview.
Start with the customer“We need an
AI claims agent.”
Leave with work you can defend.
A customer asks for an AI agent. You look at the workflow and discover something more immediate: a delivery is retried, a second task appears, and an adjuster has to untangle the mess.
This is where forward deployed engineering begins. You work alongside the customer to understand the problem, make the system useful, and help the team own it. The code matters. So do the scope, the evidence, and the person who must use it on Monday.
This guide follows one fictional insurance customer, Northstar Mutual. You’ll learn the concepts here, try a few small experiments as you read, and use the linked exercises when you are ready to build. All customer records and outcomes in the project are synthetic.
01 / UNDERSTAND THE CUSTOMER
Find the failure the customer can show you.
The claims director has promised leadership a demonstration in two weeks. The adjuster lead reports duplicate tasks. Security has not approved live policyholder data. Finance wants a measured result. Those are four different needs inside one request.
“I sometimes see two tasks for the same delivery.”
Start by asking an adjuster to show you an example. Follow it through the system. Ask what they do when the duplicate appears, what decision they still need to make, and who can accept a first repair. A useful discovery conversation ends with an agreement someone can check.
Problem: a repeated delivery can create another intake task.
First result: demonstrate safe handling of named repeat-delivery cases.
Boundary: synthetic data; adjusters retain claim decisions.
Acceptance: a named customer owner reviews the behavior and remaining risks.
Keep estimates separate from measurements. An adjuster’s guess that a duplicate takes four minutes to resolve is a useful lead. To turn it into a baseline, define when the work starts and ends, observe an appropriate sample, and record the limits.
Put it into practiceWork through customer discovery
02 / OWN THE ENGINEERING
Be able to follow a request all the way through.
Before working on the AI layer, be comfortable with a small service. Can you receive a request, validate it, write a database record, and explain the response? Can you write a failing test and use the logs to find out why it failed?
A timeout means the caller did not receive a response in time. The server may already have committed the work.
Reproduce the bad behavior with a small input. Keep a test that fails before the repair and passes afterward.
Make the rule hold when requests overlap. A database constraint and an atomic write matter more than an in-memory check.
Keep the reproduction, fix, and test understandable in a focused diff. Another engineer should be able to run your work.
If these are new to you, build a small API and a paginated data collector first. Use the engineering foundation path, then return to this project. If you already ship services, use the failure below to test your understanding.
Be ready to explain“The client timed out. How would you find out whether the write happened?”
03 / MAKE THE INTEGRATION RELIABLE
A retry should not create more work.
Northstar’s sender delivers an event. The service creates an intake task and commits it, but the response disappears. The sender retries the same event. The old service treats that repeat as another request to create a task.
Try the same delivery again. Both sides receive the same event; watch how their task counts differ.
One event. Delivered again.
The first delivery looks fine.
Record the operation with its result.
An illustration of the rule, using a browser counter. The Python exercise tests the actual service; this demo does not establish durable or concurrent correctness.
The useful concept is idempotency: repeating the same operation does not repeat its side effect. Record the operation’s identity with its payload and result, then return the recorded result for a genuine replay.
(customer identity, event ID) → payload + stored resultThat customer boundary matters. Establish customer identity from the authenticated request context rather than trusting a customer ID supplied in the payload. Harbor and Cedar can both send evt-01; they are different operations. A fresh event can legitimately update an existing claim, so the claim ID alone is not a deduplication key. If an existing event ID arrives with a different payload, surface a conflict rather than silently accepting it.
For a real service, the identity rule and write must remain correct across concurrent workers and process restarts. Explain the transaction boundary and the database constraint that enforces it. Keep tests for a lost response, a changed body, two customers sharing an ID, and overlapping retries.
Build the repairRun the Northstar Python intake exercise
Be ready to explain“Your test passes with one worker. What changes when two retries arrive together?”
04 / ADD AI YOU CAN CHECK
A model can propose. A person still needs evidence.
The duplicate-task defect can be repaired with ordinary engineering. AI becomes useful at a different boundary: proposing structured fields from claim documents so an adjuster can review them. Keep the two responsibilities clear.
A valid JSON response can still contain the wrong fact. Each proposed field needs a source reference, the relevant span, and a policy for what happens when evidence is missing or contradictory. A model’s confidence is not proof that its answer matches the document.
Loss occurred on 12 March.
Claim reported on 14 March.
Loss date: 14 March
The report date was mistaken for the loss date.
Keep customer records isolated, make approval a deliberate action, and detect a stale proposal if another reviewer changes the underlying task. Human review is a workflow with identity and version checks, not a checkbox added after the model call.
Build an evaluation set: named cases with expected behavior. Include missing fields, conflicting documents, unsupported proposals, stale versions, and cross-customer records. Inspect failures by category. A high average can hide the one failure that makes the customer unwilling to use the system.
Apply this to the projectBuild the human review workflow
Be ready to explain“The model is confident and the output is well formed. What would make you reject the proposal?”
05 / OPERATE WHAT YOU DELIVER
Make the failure understandable to the next person.
A customer sees a stalled workflow. Your job is to determine what is known, what is uncertain, and what action is safe. An error response does not always mean nothing committed. A queue becoming empty does not prove every customer received the intended result.
- Find the affected boundary. Identify the customer, operation, time window, and system that last confirmed a result.
- Reconcile before replaying. Compare delivery records with committed state. Separate known failures from unknown outcomes.
- Recover and verify. Take an authorized action, inspect the resulting state, and record what remains unresolved.
Your runbook should explain the evidence to inspect, the conditions for replay, how to disable the workflow, and who to contact. A useful handoff lets someone else make these decisions without guessing what you meant.
Practice the incidentDiagnose and recover the workflow
Then ask whether the customer actually benefits.
Adoption needs a denominator. Compare everyone eligible to use the workflow with the people who tried it and the work they completed. Look at unfinished cases and teams that stopped using it. Faster completed cases can coexist with a disappointing customer outcome.
Measure review time, rework, and support effort as well as time saved. Minutes freed are capacity; they become cash savings only when a budget line changes. The value and ROI handbook develops this measurement plan. The twelve-delivery discovery sample demonstrates a failure mechanism. It does not establish Northstar’s weekly duplicate rate or return on investment.
Test the value claimWrite a recommendation from adoption evidence
Be ready to explain“The demo works. Why might you still recommend changing or stopping the pilot?”
06 / LEAVE A TRAIL OF PROOF
Keep the work another engineer can inspect.
As you finish each part, preserve the code, the checks, and the decision it supports. A strong project story connects one customer need to a reproduction, a repair, an evaluation, a recovery procedure, and an honest outcome.
Save the exact result you are discussing. A later experiment should not silently replace the evidence attached to your earlier decision. Ask another person to review it and record the next action, not just a positive comment.
Bring the project togetherPrepare the evidence portfolio and handoff
07 / PREPARE FOR THE INTERVIEW
Explain the decision, then handle the follow-up.
Practise a two-minute account of the project: who needed it, what you personally decided, the evidence that changed your approach, and the result you can support. Describe Northstar as a synthetic project. Keep real customer experience and practice work accurately labeled.
Your follow-up practice should connect to the work you have done:
- Coding: reproduce the duplicate, repair it, and explain a test that would fail under a weaker implementation.
- Decomposition: turn “make claims faster” into a first delivery with measurable scope and clear exclusions.
- Systems: trace persistence, retries, customer isolation, latency, and recovery through the design.
- Customer judgment: respond when live access is delayed or the sponsor asks for automatic payout.
- Learning: inspect an unfamiliar dependency, test your understanding, and explain what you still need to verify.
Ask a peer to change a constraint halfway through. Defend a rejected alternative. Admit an unmeasured outcome. This is more useful practice than memorizing a polished project summary.
Rehearse with changed casesUse the interview practice cases