Skip to content

08 — Real-World Reported Questions (from online research)

Researched from public sources (early 2026). Read the credibility note first — it's the honest picture.

Credibility & the BCE-specific finding

  • No publicly reported, AI-specific interview questions exist for BCE Global Tech / Bell. Their indexed interviews (Glassdoor, Prepfully, Fishbowl) are classic QA: Selenium, Cucumber, Java/Python, API/RestAssured, CI/CD, telecom domain (subnetting, IMS, VoIP/SIP), ServiceNow. The "QA AI Specialist" title is newer than the reviews. → The AI questions below come from comparable roles, not from BCE candidates.
  • BCE process (well-corroborated): 4 rounds (2 technical → 1 managerial → 1 HR), sometimes a weekend take-home; difficulty ~3.1/5; long timelines.
  • Credibility legend: 🟢 reported from a real, named-company interview · 🟡 curated "frequently asked" blog list (plausible, not single-sourced) · 🔵 ISTQB CT-AI/CT-GenAI syllabus (the de-facto standard for this role family).

Takeaway: prepare the classic BCE QA stack + telecom domain for the technical rounds, and layer the AI-eval / RAG / red-team material from files 01–07 on top. That combination matches both the public BCE footprint and the AI-specialist title.

Highest-value REAL questions (🟢 named companies)

Question Where reported Covered in kit
"Explain how RAG works." Sierra (via Exponent) 02 Part C
"How would you evaluate an LLM feature when there's no single correct answer?" Exponent / 100-interviews 02, 06
"What does LLM-as-a-judge do, and what are its failure modes?" Exponent / Medium 02 Part B
"How do you catch regressions when you change a prompt or swap a model?" Exponent / Medium 02 Part E, practice-repo gates
"Design a RAG system for a customer-support chatbot." Multi-company common opener 05 Q1
"Design an insurance-claims RAG agent while controlling token cost." Scale AI (via Exponent) 05 Q3 + cost note below
"Design an inference batching system: throughput vs latency." Anthropic (repeated) see new note below
"A deployed model gives confident but wrong answers in high-risk contexts — investigate & mitigate." Anthropic 05 Q5, 06
"Your QA/RAG system always answers even when no answer exists — how do you detect unanswerable questions?" 100-interviews negative rejection (new section below)
"What makes a system truly agentic (vs a simple LLM chain)?" Multi-company 03 §8, 05 Q3
"Your RLHF model is gaming the reward model (reward hacking) — fix it?" 100-interviews advanced; note below
Python coding + GenAI core concepts (transformers, agentic AI) across 2 tech rounds Capgemini GenAI, India, Aug 2025 (accepted offer) 04 Part C, practice-repo

Frequently-asked (🟡 curated lists) worth drilling

  • Temperature / top-k / top-p, tokenization, context windows → 01
  • Zero/one/few-shot, chain-of-thought, when to use each → 01 (+ note below)
  • Name & explain RAGAS metrics: faithfulness, answer relevancy, context precision, context recall → 02 Part D
  • Limitations of BLEU/ROUGE/BERTScore for GenAI → 02 Part A
  • Prompt injection direct vs indirect; prevent prompt/data leakage → 03
  • LangChain vs LlamaIndex vs LangGraph → general-framework / RAG-indexing / stateful-agent-graph (note below)
  • "How would you test an ML model?" → train/test split, drift, overfitting, adversarial, A/B, metamorphic testing (note below)

Gaps this research surfaced — 5 things to ADD to your prep

The kit covers most reported questions, but these recurring themes deserve explicit answers:

1. Negative rejection / abstention ("does it say 'I don't know'?")

One of the most-repeated real questions. A RAG/QA system that always answers is dangerous. Test that when the retrieved context lacks the answer, the system abstains ("I don't have that information") instead of hallucinating.

Answer: "I add a class of golden cases that are unanswerable from the knowledge base and assert the system abstains rather than fabricates. I check grounding/faithfulness (an answer with no supporting context is a fail even if it sounds right), and I set a retrieval-confidence threshold below which the system should defer or escalate to a human. RAGAS calls the broader property noise sensitivity / negative rejection." (See the retrieval-miss case in the practice repo — faithful to wrong context but context_recall = 0.)

2. Metamorphic testing (the test-oracle problem) — ISTQB CT-AI core 🔵

When you can't know the exact correct output (no oracle), test relationships between inputs and outputs instead.

Answer: "With no fixed oracle, I use metamorphic testing: define relations that must hold under input transformations. E.g., paraphrasing a question shouldn't change the answer's meaning; translating EN↔FR should preserve the answer; adding irrelevant context shouldn't change a correct answer; reordering retrieved chunks shouldn't flip the result. A broken relation is a bug even without a gold answer." This is a strong, senior-sounding concept few candidates name — worth memorizing.

3. Token-cost / latency budgeting (production realism)

2026 interviews reward "what happens after it ships."

Answer levers: tight retrieval (fewer/smaller chunks, rerankers), model routing (cheap model for easy queries, escalate hard ones), caching, capping max_output_tokens, confidence-thresholded human fallback, and measuring cost/latency as first-class eval metrics alongside quality.

4. The framework vocabulary (so you're not caught flat)

  • LangChain — general LLM-app orchestration (chains, tools).
  • LlamaIndex — RAG/indexing-focused (ingestion, retrieval).
  • LangGraph — stateful, graph-based agent workflows (loops, branches, memory).
  • Plus the eval/red-team tools you already know: RAGAS, garak, PyRIT, promptfoo, DeepEval, Giskard.

5. Reward hacking (only if pushed on RLHF)

If they go deep on model training: a model games the reward model to get high scores without real quality.

Answer: "Harden the reward model, add a KL penalty to keep the policy near the base model, improve preference-data diversity, and evaluate against held-out behaviors the reward model never saw — so you catch gaming that the in-loop metric misses." (Advanced; don't volunteer unless asked.)

Patterns the research confirmed (what these interviews reward)

  1. Evaluation thinking is THE differentiator — measuring quality with no single right answer.
  2. Separate retrieval from generation in RAG (signals seniority).
  3. Regression testing of non-deterministic systems in CI/CD (golden set, thresholds, low temp).
  4. Negative rejection / abstention.
  5. Production realism — cost, latency, guardrails, routing, "after it ships."
  6. Red-teaming literacy is baseline, not niche.
  7. Metrics fluency — know BLEU/ROUGE/BERTScore weaknesses; name RAGAS's four metrics.
  8. ISTQB CT-AI / CT-GenAI = the canonical syllabus (metamorphic testing, oracle problem, drift).

Source quality, plainly

  • Most credible (named-company, reported): Exponent's AI-engineer guide, a "100+ real interviews" Medium compilation, a JoinTaro Capgemini-GenAI-India experience, Glassdoor candidate reports (couldn't be scraped — verify by logging in).
  • Curated/SEO (plausible, not single-sourced): testRigor, DataCamp, Analytics Vidhya lists. Treat as "commonly asked," not proof.
  • AmbitionBox / Reddit: searches surfaced no AI-QA-specific interview threads worth quoting — flagging that rather than padding.
  • BCE/Bell: only classic-QA reports exist publicly; no AI-specific BCE questions found.

→ Back to README · drill with 07-mock-interview.md