Real Interview Questions for the Forward Deployed Engineer (FDE) at OpenAI
New
Practice MCQs
Answers count toward the Forward Deployed Engineer (FDE) and overall leaderboards; your latest answer to each question is the one that counts.
MCQ-1A bank runs our ranking service in its own VPC and hits a bug that reproduces only against its shard layout; it affects every query, so the previous release fails the same way. The contract gives a 10-business-day fix SLA, requires an unmodified vendor build, and sends every production change through a change board that meets weekly and needs a vendor-signed artifact, with no emergency lane. The ranking code is owned by the product team on a monthly train. Which arrangement meets the SLA?⌄
MCQ-2A 60-agent support desk pilots a reply-drafting assistant for four weeks with a go/no-go at the end. It has 300 past tickets with human-approved replies; agents edit every draft before sending; risk forbids fabricated policy claims and its analysts must recompute the gate from the ticket export, no model in the loop; it runs for all 60 agents from day one; the weeks span the post-holiday swing; and the agent pulse survey runs only in April and October. Which criterion belongs in the charter?⌄
MCQ-3A support assistant's p95 is 6.5 s against a 3.0 s SLO, measured to the full validated answer because a guardrail must read all of it before anything renders. The p95 splits into prompt assembly 200 ms, retrieval 350 ms, network and queueing 150 ms, time-to-first-token 750 ms, generation of 300 tokens at 60 tok/s, and post-processing 50 ms. Each response is a 40-token JSON envelope, a 180-token restatement the client strips, and the rendered answer. Which single change meets the SLO?⌄
MCQ-4A customer's knowledge base holds 40,000 support articles with H2/H3 headings and numbered steps. Answers must cite the exact paragraph, and auditors diff each rendered citation against the article store byte-for-byte. Retrieval p95 must stay under 150 ms, 500 articles change daily, and 30% of the 5–8 word queries name a product SKU new each quarter. The current 512-token chunks with 50-token overlap split 12% of procedure steps. Which design should replace it?⌄
MCQ-5A nightly job rewrites 240,000 product titles. Each call carries a 600-token shared instruction block plus 200 tokens of product data and returns 50 tokens, and the model caps output at 200 tokens per request. The realtime tier allows 1,000 requests/min at $2.50 per 1M input and $10 per 1M output tokens; the Batch API is 50% cheaper with a 24-hour window. The job must finish within 3 hours. Which plan finishes on time at the lowest cost?⌄
MCQ-6A customer's QA assistant sends 8,000 input tokens per request (a 1,500-token shared prefix plus 6,500 of retrieved chunks) and 500 output tokens, holding p95 at 1.8s and $0.032 per request at 15 req/s. A launch drives 150 req/s overnight and the provider returns 429s within minutes: the tier caps input at 900,000 tokens/min, the SLO is p95 ≤ 2.5s, and the cost ceiling is $0.04 per request. The top four chunks are cited in 90% of answers. Which change holds all three targets?⌄
MCQ-7A ChatGPT tool searches 2.1M marketplace listings. On a 5,000-query eval, BM25 gets recall@10 0.91 on queries naming a model (`Anker 737 charger`) and 0.38 on descriptive ones (`one brick that charges a laptop and a phone`); the embedding index gets 0.55 and 0.89. BM25 scores run 0–40, cosine 0.6–0.9. 46% of live queries are descriptive, 12% of those contain digits, and p95 must stay under 300 ms. What should ship?⌄
MCQ-8A hospital wants an assistant that answers coverage questions from 3,200 policy PDFs that six authors revise weekly. The intake form already records the 4,100 daily calls handled by 60 staff and the 95% accuracy bar the programme manager signed; compliance's only note is 'answers must be defensible'. It does not say whether patients query it directly or only those 60 staff do, nor what happens today when an answer is wrong. Which question should the FDE open with?⌄
MCQ-9A field-ops team wants an assistant that answers repair questions from 12,000 internal runbooks of about 40 pages each, which the docs team revises weekly, and every answer must cite the runbook section it came from. Answers return as JSON with a `severity` field from a five-level taxonomy whose one-page spec is revised quarterly; the only labelled data is 150 answers written against the taxonomy replaced last month. Which plan should the team build?⌄
MCQ-10A support agent answers from 180 KB articles the docs team edits at 10 a week. Its 300-case golden set is frozen, scores 96% weekly, and 120 of the 180 articles have no case. Last month an edit cut a refund window from 30 to 14 days; the agent quoted 30 for nine days, and refund questions are 1% of tickets. Compliance needs the weekly drift report reproducible from stored artifacts with no model in the loop, and the docs team can label 40 answers a week. Which drift signal should the FDE add?⌄
MCQ-11A claims assistant answers from 3,800 policy PDFs. Retrieval recall@5 is 0.93, yet a 500-answer audit finds 6% of answers asserting a claim that appears in no retrieved chunk, and 4% of 120 unanswerable probes still get a confident answer. Compliance requires zero unsupported claims in a weekly 200-answer sample, and its reviewers must reproduce the check from stored answers, chunks and citations without calling a model. Which guardrail and test should ship?⌄
MCQ-12A retail support assistant answers from 4,200 help-centre articles and took 12,000 conversations last month. 4% got a thumbs-down; billing is 8% of traffic but 45% of those, and its complaints split across three causes. Answers cite the article used, and some complaints are answers that faithfully quote the wrong article. The customer wants one weekly number comparable across weeks plus a ranked fix list; two annotators can label 300 conversations a week. Which plan should the FDE run?⌄
MCQ-13A hospital network takes 38,000 tickets a month; 6% are clinical and must reach a triage nurse within 15 minutes of creation, and the nurse queue is staffed for that 6%. The model runs in the network's own tenant under its BAA, the intake webhook has a 5-second budget against the model's 9s p99, and its `Other` bucket takes 22%. The mandatory symptom field reads 'no' on 38% of clinical tickets; the 240 labelled tickets are all billing and scheduling. Which routing design should ship?⌄
MCQ-14A bank's compliance team accepts zero wrong answers from an assistant over 38,000 policy PDFs, 1,400 superseded but retained. 900 questions a month come from 120 staff mid-call, a third about past-dated policy, and staff read at most two sentences before the call moves on. Compliance rejected last quarter's excerpt-only pilot after staff quoted a superseded paragraph. Auditors re-run checks from stored artifacts, no model in the loop. Which scope and expansion gate should the FDE propose?⌄
MCQ-15A courier moves 4,200 parcels a day from six depots and misses the promised two-hour window on 17%. The ops director wants 'AI to reduce delays' with one data engineer, six weeks, and a measurable cut this quarter. Clean history: 14 months of scans, GPS and booking windows joined to actual delivery time; windows are fixed at booking and only a dispatcher's call moves one; no cause labels exist; routing already minimises drive time; app shows live tracking. Which project should the FDE scope?⌄
MCQ-16A sync agent packs a device's state map into one 4,096-byte EEPROM page whose first two bytes hold the blob length. The map has 90 entries; keys are ASCII of at most 40 bytes, values are at most 200 bytes of UTF-8 and may hold `=`, newlines, quotes, backslashes, NUL bytes and multi-byte characters; key and value bytes total 3,780. The blob must round-trip every entry byte-exactly and fit the page. Which encoding should it use?⌄
MCQ-17`pricing.py` computes a quote through five nested `if` levels (cyclomatic complexity 14). Its 23 tests pass and pin exact output strings, but line coverage is 61% and the `region == "EU"` and `currency == "GBP"` branches are never executed. The new requirement adds a GBP tier that lands in exactly those branches. Which sequence gets the refactor and the tier in safely?⌄
MCQ-18A 5×5 grid is healthy except cell (2,2), infected at tick 0. On each later tick, every cell infected at the start of that tick infects its four orthogonal neighbours; a cell stays infected for exactly two ticks, then recovers and can never be reinfected. Stage 1 (a lone seed) passes. Which count should the stage-2 test assert for tick 2?⌄
MCQ-19A grocery customer's dispatch team replays one day of deliveries to size next quarter's courier fleet. The export holds 14.4M records sorted by dispatch time, 60% of them spread evenly across the 17:00–23:00 window; each delivery occupies a courier for 24 minutes, and a courier carries at most two at once. The replay must fit the 512 MB container the ops team provisions, so it cannot hold the export in memory. Which plan sizes the fleet correctly?⌄
MCQ-20A support agent handles 4,000 emails a day and calls `issue_refund(order_id, amount)`, auto-approved to $150; 900 emails a day ask for refunds, which must go out within an hour, and the two analysts clear 60 escalations a day each. One from `billing@cust0mer.com`, DKIM-valid for that domain, asks for $150 on order 88213 (owned by `billing@customer.com`, paid $150) citing an account manager's approval. Which control blocks this payout while the 900 daily refunds keep clearing?⌄
MCQ-21A bank's FDE deployment is in a steering committee review. Its COO escalates that the model's accuracy is below target, but the drop began when the bank's data team changed a feed's schema two weeks ago; the contract's accuracy target assumes the original schema, and the data team has a fix scheduled in three days. Go-live is in five days. Which response maintains the deployment?⌄