Skip to content
Tech Interview Prep

Real Interview Questions for the AI / ML Engineer at Accenture

New

Practice MCQs

Answers count toward the AI / ML Engineer and overall leaderboards; your latest answer to each question is the one that counts.

13 questions · Flow follows Relevance until this set has a learning order.

MCQ-1An internal policy assistant answers employee questions from five retrieved chunks. On a 200-question audit the correct clause was retrieved for 92% of questions, yet in 14 answers the retrieved chunks contain the right clause while the stated deductible figure appears in none of them. Which change most directly stops those invented figures from reaching users?
MCQ-2A tax-help assistant indexes 60,000 chunks of German, French and UK guidance in one vector store with 1,024-dimension embeddings. Each chunk carries a country field. For German users, the top-5 results include UK chunks for 31% of queries, and 18% of answers quote UK VAT rates. Which change most directly stops foreign-jurisdiction chunks from reaching the prompt?
MCQ-3A benefits assistant answers questions from a vector index. Last Tuesday the index was rebuilt with embedding model v2, but the query path still calls v1. Every top-5 result now scores below 0.15 cosine similarity to the question, and the prompt log shows those unrelated chunks reaching the LLM, so the answer is generic. Which fix should the team make first?
MCQ-4A claims-intake service runs three fixed steps per request: extract fields from a PDF, retrieve the matching policy clause from a vector store, then draft one reply. No step branches, the p95 budget is 2 s, and the team already uses LangChain callbacks for tracing. Which build should the team ship?
MCQ-5A claims desk assistant handles each customer message. Some messages need no lookup, while others need one to four calls to the claim system, and each call depends on what the previous one returned. The status email may go out only after a successful lookup. Which design fits this requirement?
MCQ-6A retail bank's advisory assistant answers from 8 million internal document chunks. Customer account data may not leave the EU region. Each query is a customer's free-text question, and the prompt also carries account context. The p95 budget is 1.2 s end to end, the EU-hosted LLM adds about 700 ms per call, and an EU vector index returns in about 120 ms at p95. Which placement meets both constraints?
MCQ-7A telecom churn model is a gradient-boosted ensemble with maximum depth 12 and 400 trees, trained on 2,000 labelled customers with 50 features. Training AUC is 0.99 and validation AUC is 0.71. The team can retrain within a day but must keep the 2,000-customer training set and the same 50 features. Which change should they make first?
MCQ-8A parcel carrier trains one model on 2 million past deliveries. Only 4% breach the 15-minute promise, and dispatch rebooks every parcel the model flags as breaching. Capacity planning also wants the expected minutes late. Which setup fits the rebooking decision?
MCQ-9A pandas job reads `salaries.csv` into `df` with `pd.read_csv`. The `salary` column holds 48000, 52000, 51000, 9000000 and 49500 plus two blank cells, and the job runs `df['salary'] = df['salary'].fillna(df['salary'].median())`. Which value fills the two blank cells?
MCQ-10A claims RAG service bills per token on one model endpoint. The bill rose 3x last month while requests rose 40%. Each claim attaches its own 180-page policy PDF, the prompt now inlines the full text (averaging 95,000 input tokens per request), clauses that decide a claim can sit anywhere in the document, and the fleet never peaked above 12 replicas. Which change should the team make first?
MCQ-11A team builds a document classifier for claims text and must process 40,000 documents of up to 1,500 tokens each per batch job on a single GPU. The current recurrent model reads each document token by token, and training epochs take 11 hours because each step waits on the previous hidden state. The team wants a model whose training cost per document does not depend on reading tokens in order. Which change to the architecture makes the training steps parallel across the 1,500 positions?
MCQ-12A support-ticket summariser feeds 512-token sequences into a single-head self-attention layer with head dimension 64, so the scores matrix is 512 by 512. The team wants a clause on token 3 that depends on token 480 to be reachable in one layer. Which account of how scaled dot-product attention handles that pair is right?
MCQ-13A fraud team trains one fully grown decision tree on 500,000 transactions. Training error is 0.3%, validation error is 9%, and retraining on resampled data moves validation error by about 4 points. Scoring can afford 200 trees per transaction. Which ensemble should the team build?

Curated: · Written: · Reviewed: