Rubric-graded simulations for engineers who can build AI demos and want to deploy them in front of customers. You'll practice the CEO pitch, the discovery call and the 3am incident, and get scored on each.
Who it's for: aspiring FDEs and the solutions, customer and implementation engineers doing the same work. The bar is Python and one real LLM project.
Try a graded drill with no account. Sign in free to unlock Modules 1 to 3. All-Access has a 30-day money-back guarantee.
M01 · Boss · Stage 1 of 3Live grader
CEO
The Skeptical CEO
Budget owner at an insurance company · simulated stakeholder
Presses on value first, then reliability, then invented numbers.
You're pitching PolicyPilot, an AI reply assistant for the insurance help desk.
View brief
The help desk has 40 agents.
Replies use approved policy documents and cite the relevant clause.
A human agent approves every draft.
The assistant never decides claims or coverage.
Agents currently spend about 20 minutes per ticket searching policy documents.
CEO
The CEO: I have 60 seconds. What does it do, and why should I fund it? One paragraph. No jargon.
Rubric: The One-Paragraph BriefPass mark 60 / 100Stages 2-3 in the app
Your answer
rubric score
evidence
the CEO's reaction
revise
The three moments a Forward Deployed Engineer is judged on.
Each leaves a document behind, here for the same help-desk deployment the CEO is asking about.
In the boardroom
Module 1 · the Skeptical CEO
Common failure mode
Explaining architecture when the executive asked for a decision.
FDE standard
Value, boundaries and risk in under 60 seconds.
Output
A one-paragraph executive brief.
Executive brief: PolicyPilotto: CEO · 1 paragraph
Decision requested
Fund a four-week pilot on one ticket category
Business impact
40 agents spend about 20 minutes per ticket searching policy documents
Human approval
Required. An agent approves every draft.
Out of scope
Claims and coverage decisions
Risk posture
No approved clause, no answer: it says so instead of guessing
Next question in the room, stage 2: “Your demo worked once. What happens on the 500th weird customer message, and how do you know?”
In discovery
Module 3 · a live call with Meridian Mutual's CTO, compliance and claims ops
Common failure mode
Taking the feature request literally and scoping a six-month build.
FDE standard
Find the real constraint; cut the smallest slice that can ship.
Output
A scoped first slice with its assumptions and exclusions.
Discovery notes: help deskscoping · slice 1
Customer ask
“We need an AI chatbot.” A feature, not a problem.
↓
Actual bottleneck
Agents spend about 20 minutes per ticket finding the right policy clause.
↓
Slice 1
Draft the reply and cite the clause
An agent approves every draft
Measure handle time on one ticket category
Out of scope
Claims decisions
Replies sent without an agent
Interpreting ambiguous policy
Assumptions to confirm
The policy library is current and approved. The help desk can tag tickets by category.
In production
Modules 4 and 5 · ending in the 3am page
Common failure mode
Treating the prototype as the deployment.
FDE standard
Least privilege, traces, an eval gate, cost alerts and a rehearsed rollback.
Output
A readiness review, and an incident call you can run.
Day-2 readiness review: PolicyPilot6 of 7 checks
✓Passed: Least-privilege service account2 roles, no project-wide accessM05
✓Passed: Secrets out of code and configread from a secret storeM05
✓Passed: Traces across retrieval, model, toolp95, errors, tokens per requestM05
✓Passed: Golden eval set reruns every releaserelease stops if misses riseM05
✓Passed: An agent approves every draftno reply leaves without a humanM01
✓Passed: Rollback to last good revisionrehearsed, under five minutesM05
!Missing: Alert on cost per requestlatency alerted, spend is notM05
DecisionNO-GO until the cost alert exists
See exactly how an answer is judged.
Two real attempts at the Module 1 boss, measured and scored by the same pipeline that scores yours. Unedited; only the highlights are ours.
“But what happens when it gets a clause wrong, and how do we safely test this first?”
The simulated CEO, after attempt 1 scored 15
Submission
Evidence
Rubric
Reaction
Retry
Record
M01 / The Skeptical CEO / Stage 1 of 3: The One-Paragraph Brief
Submission
PolicyPilot is a RAG pipeline on Gemini: we chunk your policy PDFs, embed them into a vector store, retrieve the top-k clauses for each ticket and generate a grounded draft with citations. It plugs into the help desk through an API, and we can fine-tune the prompts as we learn. It should make the support team a lot faster.
Highlighted: the phrases the evaluator's reasons point to.
Evidence · measured by code
✓Present: Within 120 words, one paragraph
✗Missing: Plain language
✗Missing: Business problem stated
✓Present: Business outcome stated
!Check: Value backed by a number
✗Missing: Human approval point stated
✗Missing: Limit or failure behaviour stated
✗Missing: Way to prove it before scaling
!Check: Opens with the architecture, not the business
Graded 5 Oct 2026 · 59 words · pass mark 60
Rubric · scored by the evaluator
15/100NOT YET
Executive clarity5 / 25
The pitch opens with a technical architecture tour filled with jargon like RAG, vector stores, and chunking rather than focusing on executive outcomes.
Value framing8 / 25
It mentions making the support team faster, but omits the current time spent searching policies and lacks any quantified target.
Boundaries2 / 25
It fails to state that human agents approve every draft or that the assistant never makes coverage decisions.
Risk and proof0 / 25
There is no mention of how the system handles uncertainty or any pilot decision gate to prove value before scaling.
Feedback
CEO
The CEO: “I get that you want to speed up support tickets using our policy docs. But what happens when it gets a clause wrong, and how do we safely test this first?”
Retry
What would improve this answer →
Reframe the pitch around the 20 minutes agents spend searching policy documents per ticket, ensuring every draft requires human approval before sending.
Replace technical terms like RAG, vector store, and top-k with plain business language.
State the baseline operational problem, human approval guardrails, and a pilot proof plan.
What landed: Keeps the submission short and well within the word limit.
Submission
Our 40 help-desk agents spend about 20 minutes per ticket hunting through policy documents. PolicyPilot finds the approved clause and drafts the reply with that clause quoted, and an agent approves every draft before it goes out. It never decides claims or coverage, and when the policy doesn't cover a question it says so instead of guessing. I'm asking you to fund a four-week pilot on one ticket category: if handle time drops and accuracy holds, we expand; if not, we stop.
Highlighted: what the evaluator credited.
Evidence · measured by code
✓Present: Within 120 words, one paragraph
✓Present: Plain language
✓Present: Business problem stated
✓Present: Business outcome stated
✓Present: Value backed by a number
✓Present: Human approval point stated
✓Present: Limit or failure behaviour stated
✓Present: Way to prove it before scaling
Graded 5 Oct 2026 · 82 words · pass mark 60
Rubric · scored by the evaluator
99/100PASSED
Up 84 from your last try (15 → 99).
Executive clarity25 / 25
The pitch is concise, free of jargon, and opens directly with the business problem and practical solution in a single paragraph.
Value framing24 / 25
It uses factual baselines from the brief regarding team size and ticket search time, establishing a clear handle time target without making up speculative ROI figures.
Boundaries25 / 25
It establishes clear guardrails by requiring human approval on every draft and explicitly stating that the system never decides claims or coverage.
Risk and proof25 / 25
It outlines safe behavior when policies are missing and proposes a structured four-week pilot with clear continuation criteria.
Feedback
CEO
The CEO: “I get it: human-in-the-loop drafting to shave time off policy search without making claims decisions. Show me the exact target metrics for this pilot and let's get started.”
Retry
To push it further →
Include a specific target handle-time reduction figure for the pilot to quantify success criteria.
Specify how accuracy will be measured alongside handle time during the four-week test.
Stages 2 and 3, reliability and trust, are in the app. Sign in free to unlock Modules 1 to 3.
Nine prep lessons, then six graded mock rounds, from ownership screen to client interview
You'll be able to: Walk into each FDE interview round already graded on it: ownership, decomposition, coding, learning, AI solution design and the client interview.
Train, retry and revisit every live simulation for 12 months.
53lessons available now
6boss battles, each graded against a written rubric
6mock interview rounds
12months of access, renewing yearly until you cancel
Module 6 is included now; the capstone exam is included when it ships. Access starts when payment completes, on your FDE Spot account (or the one you create later with the email you pay with). Billed yearly through Stripe; email [email protected] to cancel. 30-day money-back guarantee. Recently laid off? $49/yr, no conditions: email us.
In development, no ship dates: the Module 7 full-loop boss, the capstone exam and the FDE credential.
Enterprise, for teams
Custom quote
Quotes start at $299. Seats, billing period and total price are set out in a written quote before anything is charged.
Free drill, no account. Try a graded drill with no account. Sign in free, with Google or any email address, to unlock Modules 1 to 3 and their boss battles, plus preview lessons in Modules 4, 5 and 7.
Software engineers who can already build with LLMs and want the customer-facing half of the FDE job: discovery, scoping, production readiness and the interview loop. The bar is Python and one real LLM project behind you; no Google Cloud experience is assumed. If you haven't built anything with an LLM yet, this isn't the place to start. If you have shipped an AI demo and want to be the person a customer trusts with the deployment, start with Module 1.
A year of access, for one learner, to everything live: 53 lessons, 6 boss battles and 6 mock interview rounds. Module 6 is included now; the capstone exam is added when it ships. It's access, not a certificate.
Browser-based, nothing to install. Short reading lessons, Python exercises whose tests run in your browser, and boss battles where you answer a simulated stakeholder in timed stages. Boss stages are graded by AI against written rubrics; code exercises are checked by their tests.
Module 7 prepares you for FDE interviews: nine prep lessons and six graded mock rounds, from ownership screen to client interview.
Yes. Try a graded drill with no account: Stage 1 of the Module 1 boss battle is graded live at the top of this page, with your answer, our rubric and a real score, and no card. A few attempts a day per visitor, because each one is a real model call we pay for. The first lesson is also readable with no account. Sign in free, with Google or with any email address and a password, to unlock Modules 1 to 3.
For the Module 1 CEO brief, code first measures your text: length, technical jargon, and whether you state the problem, the outcome, a human approval point, a limit and a way to prove it. An AI evaluator then scores four rubric dimensions out of 25 each, with a reason for every score, against a pass mark of 60, and the simulated CEO reacts separately; the reaction never changes the score. You see all of it, revise and send again. Other stages are scored by the same kind of AI grader against their written rubric. Two real graded attempts are shown above. There is no human marking; human help is by email.
No. FDE Spot is an independent platform. We use Gemini as our primary AI stack, but we're not endorsed by Google. Think of it like how AWS study guides aren't made by Amazon.
Modules 1-6, the Module 7 interview loop (9 prep lessons and 6 mock-interview rounds), and Module 8's agent security labs are live. Try a graded drill with no account. Sign in free to unlock Modules 1 to 3. Modules 4 and up come with the All-Access pass ($99/yr for one learner), and each keeps free preview lessons. The Module 7 full-loop boss, the capstone exam and the credential are in development, with no ship dates promised. Teams can buy Enterprise by enquiry. Founding-cohort seats already granted keep their access, including Module 6.
Your annual subscription gives you 12 months of access to all eight live modules. We don't quote a single hour count, because the open-ended briefs take as long as your answers do. To size it: a reading lesson is 3-7 short pages, and boss stages have a five-minute clock. There are no extra costs: code runs in your browser and grading runs on our side, so you need no API key or cloud account. Feedback comes from AI grading against written rubrics; human help is by email.
We offer a 30-day money-back guarantee, counted from the day you pay: email us inside that window and we refund, no questions. It does not wait for the next module to ship. Try a graded drill with no account, and read the first lesson too, before deciding. Sign in free to unlock Modules 1 to 3. If you're an EU consumer, this is in addition to, not instead of, your statutory 14-day withdrawal right, which runs from the day after purchase.
