Interview practice

Forward-deployed engineer and forward operator

Two role shapes you can practice from the six weeks and the five fictional cases. This page is not a certificate and not a job offer from this site.

Posted experiences from other people are a separate page: Interview kit. They are not merged into the answers below.

What the two titles usually mean

Teams use the words differently. This is the shape to practice, not a hiring promise.

• Forward-deployed engineer. You build the pilot and you sit with the customer. You write the discovery note, the metric in their words, the scope, the written no, the field map, the corpus list, and the evals. You refuse the tool that their login cannot safely hold.
• Forward operator. You run the live pilot after it is up. You watch the failures, you keep the old screen during a parallel run, you write the five-line incident note, and you write the handover their on-call can follow without you.

One person sometimes does both on a short pilot. The split still matters in an interview. Building the loop is not the same work as noticing that a green health check hid a bad reply.

A day in the work

You can walk this list in plain language. Each line is already somewhere in the studio.

• Discovery. Who does the work today, not only who signed the ask. Week 1. Harbor Lane's note names the front desk, not only Nia.
• The buyer's metric. Their sentence, plus a way to check it. A deflection rate is the share of questions that never reach a person. It is a bad metric when the answer can be wrong. Week 1.
• Scope. What is in, what is out, and the date or the condition that ends the pilot. An empty out-list means every new ask is still "phase 1."
• Saying no. One refusal in writing. Harbor Lane will not send and will not read billing. Ridge Route will not connect a write on an admin login.
• The field map. Their name, your name, the source of truth, and the sentence you return when the sides disagree. Week 2. Keel and Bin: the nightly file is not the live count.
• Data classification. Who may read a document, what the corpus is, and what stays out. Week 3. Mill City leaves pay out. Indexing the whole folder is a named wrong call.
• Evals. You score the action, not the tone. A billing question must not return billing. A price question must not return a price. Week 4.
• The demo account. An account created for the pilot. Not your login. Not their admin password. A bearer is the shared secret that route requires. Week 5.
• The runbook. How it is down, how to stop it, the previous version, and who may run those steps. Week 5.
• The incident note. Five lines: what broke, who is affected, what you stopped, what you do not know, and when the next note comes. A person sends it. Week 6. Lumen Desk, ticket LD-18.
• Adoption versus page views. Count the action the buyer cared about. Opening a page is not that action. A green /health does not turn off the old screen. Keep the parallel run. Week 6.

Questions, with a short model answer

Each answer uses only this studio's fictional practice. Say that before you tell the story.

1. Tell me about a pilot you scoped, and what you refused.

I studied the Harbor Lane Clinic case. It is fictional. The front desk typed replies about hours and about moving a visit. The buyer, Nia Okonkwo, wanted a draft waiting in a staff queue before open. I would put sending, billing, clinical notes, and chart writes on the out-list. The refusal was a send button, and a second refusal was reading billing. The metric was the draft in the queue for those two message types, not a deflection rate. A person still sends the mail.

2. How did you handle a nightly file that was not live data?

I studied Keel and Bin Parts. The nightly export is a snapshot with a time on it. The live counter is the source of truth. On the field map their name is "on-hand" and the tool's name is live_qty. If the file and the counter disagree, or if the live read does not answer, the picker is told the live count did not answer. The file is not quoted as if it were now. Part KB-2042 is only a label in that story. I would not invent a change in stockouts.

3. Who was allowed to read the documents?

I studied Mill City Permits. Ruth Ellison owned the list. In the pilot index: the published guide pages she named. Out: applicant files, draft decisions, fee and pay pages, and any other folder. Who may read is filtered before text is ranked, not after the model has already seen it. Indexing the whole drive was the wrong call, and it was refused. A fee question gets "outside this corpus," not an amount.

4. What did you do when the only login was admin?

I studied Ridge Route Field. Andre Pell wanted the tool to close a job, and the only login offered was his admin account. I would not connect the write tool. I would not demo with that password or with my own admin login. A bearer is the shared secret the route requires, and an admin bearer does not belong in a field app. The pilot stops until someone creates a technician account that can write only its own note. The handover is that stop note. I would not say any job was closed.

5. What did you watch after deploy?

I studied Lumen Desk. After the new path was up, I would watch replies a person edited and sent, next to replies still sent on the old screen. I would not watch page views. I would not treat a green /health as a reason to turn the old screen off. The parallel run stays. For that shop, traces were kept seven days because someone asked how long they should live and the practice note used seven. That number is not a universal rule. Emails stay out of the trace.

6. How would you explain a failure in five lines?

On Lumen Desk, ticket LD-18 is the practice failure. The new path showed a status the old screen did not. The draft has five lines: what broke, who is affected, what you stopped, what you do not know, and when the next note comes. I would not write "no customer impact" because health was green. I would not send the note myself in the story. A person sends it. The old screen stays on.

7. What belongs in the discovery note before you build?

Who does the work, the metric in the buyer's words and in a sentence you can check, what is in and out, and one written no. On Harbor Lane the checkable sentence is that a draft exists for hours and reschedule messages and a person still sends it. The no is send, and billing. If I delete the model from that note, the desk's morning work is still described. That is the test week 1 asks for.

8. Two systems disagree. What do you tell the person?

I do not average them and I do not copy the stale side. At Keel and Bin I say the live count did not answer. At Harbor Lane, if the appointment book is silent, the draft says the book did not answer. It does not reuse a time the sender pasted into the mail. Week 2's rule is the same on the lab's orders: a disagreement tells the shopper the order system did not answer.

9. Why leave pay out of the index?

Pay is a different class from the pages the pilot was allowed to answer. At Mill City the fee table can sit in the same folder as the guide and still be out. The buyer did not ask the pilot to quote fees. A question about a fee is outside the corpus. Leaving pay out is both scope and classification. Week 3 says the same thing about the lab: leave pay out. Indexing the folder would pull it in. That is why the folder is the wrong unit.

10. What is in the handover when a read-only pilot does ship?

How it is down, how to stop it, the previous version tag, who may run those steps, and the pilot account, which is not my login. Harbor Lane hands over a draft queue and keeps staff sending mail themselves. Lumen Desk hands over a parallel run, the seven-day trace note, and an adoption table that is allowed to say stop. Ridge Route does not get this runbook, because it did not ship. Its handover is the refusal.

11. What will you not claim about this studio?

I studied the notes and five fictional engagements: Harbor Lane Clinic, Keel and Bin Parts, Mill City Permits, Ridge Route Field, and Lumen Desk. I was not placed at any of them. Northline Orders is the lab, with orders 1042, 1043, and 1044. Thirty-one checks in python3 test_all.py means the lab files still match the rules. It is not a credential, not an interview pass, and not a score. Mark as complete is local to the browser. I do not say I am certified. This site does not offer a certificate or a job.

Self-check rubric, 0 to 3

Use this on your own answers to the questions above, or ask a practice partner to use it. 0 is missing, 1 is partial, 2 is acceptable, 3 is production-minded. It is practice feedback for you. A 3 is not a job credential, not a score this site records, and not something to put on a resume. Portfolio policy.

If you score yourself a 3 everywhere, say "I studied this fictional engagement" in the same breath. The rubric measures how you explain practice, not what you have shipped.

Honest resume lines

These are lines you can put on a resume. They stay inside the fictional frame.

• Built and tested a fictional order agent; 31 local checks; not a deployment.
• Studied five fictional engagements and wrote the refusals. Not a placement and not a certificate.

A 30-minute mock

Set a timer for 30 minutes. Read this prompt once, then answer it without the page in front of you.

You have 30 minutes. Pick forward-deployed or forward operator. Using only fictional practice, walk one case: who does the work, the metric in their words, what you refused, and how you would test the refusal. Then point at the Northline lab: orders 1042 shipped, 1043 paid, and 1044 cancelled, and one control from the six weeks. End with what you will not claim. Say "I studied this." Do not say "I was placed."

A stakeholder who wants the send button

I studied Harbor Lane Clinic. It is fictional. Nia asked to connect send so the desk would save a click. I would say: the queue is the product, a person still sends, and there is no send tool. The metric is a draft waiting before open, not a message that left by itself. I would not add the button to be helpful. I was not placed at the clinic.

A stakeholder who wants billing in the index

I studied the same fictional clinic, and the same shape at Mill City Permits, where the ask was pay. I would say: billing is a different class, the folder is not the corpus, and indexing the whole drive was the wrong call. A billing question gets no balance, no code, and no claim number. I studied this. I was not placed there.

What not to claim

• Do not say you were placed, staffed, or deployed at Harbor Lane, Keel and Bin, Mill City, Ridge Route, Lumen Desk, or Northline.
• Do not say you are certified, or that a lab pass is a credential.
• Do not invent a scorecard, a rating, a ticket drop, or a savings number. None of these pages measured one.
• Do not call a green health check proof of no customer impact.
• Do not describe Mark as complete as a result anyone else can see.

How to talk about these cases

Start with "I studied this fictional engagement." Then the buyer, the metric they named, what you refused, and how you would test the refusal. Stop before any sentence that sounds like you worked there. "I deployed this at the company" is the sentence to avoid. The five cases and the six weeks are enough material. You do not need a fake client list.

Pro practice next to this page

• 45-minute coding practice. Parse a line, join it to the three orders, write once, and test that. No puzzles. Not part of the 31 checks.
• Prompt, retrieval, or fine-tune. The default is do not fine-tune.
• Keel and Bin reconciliation. Required practice for Week 2 and the Keel and Bin case. The nightly file is still not the live count.
• Harbor Lane, consulting and life-sciences stream. Written no, corpus exclusion, parallel run, handover.
• Identity and audit on week 3, and the week 6 cost sheet.
• Optional typed tools and Text2SQL. Not core.

The five cases

• Harbor Lane Clinic. Drafts only. No send. No billing.
• Keel and Bin Parts. The nightly file is not the live count.
• Mill City Permits. A named corpus. Pay stays out.
• Ridge Route Field. Admin was the only login, so the write stayed off.
• Lumen Desk. Parallel run, five-line note, old screen stays.