Interview 100 Questions for AI Engineers

A categorized set of 100 questions with clear answers, three examples each from the fca project context, simple English, and an ASCII flow chart per topic. Use these to interview prep, build flashcards, or run weekly reviews.

How to use

• Pick one category per day; read the question and answer first.
• Then read the three examples and the data-flow chart; close the page and try to retell the chart in your own words.
• End of week: run the questions for the categories you covered against a colleague or a LLM judge (use the LLM-as-judge prompt from the prompt-engineering-fundamentals section).
• Update or replace weak answers; the back-of-the-envelope math in the answer must stay correct.

Topics by category

Backend & API

• What is FastAPI Depends() and how does it enable clean RBAC wiring?- How does JWT + bcrypt together protect an API?- What is a LangGraph state machine and why prefer it over a linear chain?- How do you run async-only Python from a synchronous Celery worker?- Why use a supervisor agent instead of letting LLM agents debate freely?- How does LangGraph pause safely before a dangerous action?- How does Server-Sent Events stream a LangGraph run to a browser?- Why use Pydantic’s BaseSettings instead of os.environ dictionaries?- How do you combine IDOR checks with OAuth2 scopes on a single FastAPI route?- Why pipe logs through a ContextVar instead of passing a logger everywhere?

Database & RAG

• How do you enable pgvector and align the column type with the embedding model?- Why split documents into chunks before embedding, and what splits lose precision?- What’s the right async pool config for test vs prod?- Why does near-duplicate detection matter for RAG retrieval?- How do you confirm your HNSW index is actually being used?- Walk through an end-to-end RAG pipeline at a high level.- Why is Recall@K the right offline metric for RAG?- NullPool vs default pool in test/dev/prod — when do you pick each?- How do you detect a batch-of-vectors with the wrong dimension before it poisons the index?- When do you pick HNSW vs IVFFlat for a vector column?

DevOps & Deployment

• Why split a Dockerfile into multiple stages?- How do you orchestrate pgvector + FastAPI + Celery + Streamlit in dev?- What should a Python FastAPI GitHub Actions pipeline enforce?- What’s the difference between liveness, readiness, and startup probes?- What goes into pyproject.toml for a Python AI project?- Why commit per-resource during a big database seed?- How do you make pytest work async + use a fresh DB per run?- When do you pick HPA vs cluster-autoscaler?- Why pin CPU-only torch first in a Dockerfile?- How does bind-mounting the source folder speed up dev?

Security & Compliance

• What does Presidio do and where in the pipeline should it sit?- How does Lakera detect prompt-injection attempts?- When is HS256 right, when do you need RS256?- How do you build role-and-scope RBAC with FastAPI Security?- Why pick bcrypt over Argon2id (or vice versa)?- Why keep an authoritative list of compliance violations in code?- How do you structurally prevent IDOR beyond route-level checks?- When is a secret manager meaningfully better than env vars?- What does layered prompt-injection defense look like end-to-end?- How should you handle PII end-to-end through an LLM app?

Observability & Tracing

• How do you instrument FastAPI with Prometheus?- What does Langfuse add on top of your logger?- Why JSON logs beat plain logs at scale?- How do you add a custom business metric alongside the defaults?- Why have one custom exception base + per-domain subclasses?- How do you stream structured logs to a browser alongside agent events?- How do you define an SLI and an SLO that you can actually measure?- Why track token usage per request in metrics?- When do you pick OpenLIT over Langfuse?- Why rerun an eval on production traces that look bad?

DSA Patterns

• When does the two-pointer pattern beat hash-map lookups?- When is the sliding window the right pattern?- When does BFS beat Dijkstra?- How does binary search on the answer work?- How do you spot a DP problem quickly?- When does a monotonic stack beat sorting?- When is Union-Find the right pattern?- When does a trie beat a hash table for prefix lookups?- What’s the universal three-step skeleton for backtracking?- When is a bitmask the right representation for subsets?

System Design

• When is Cache-Aside better than Read-Through?- How does a per-user token bucket beat a fixed-window counter?- When is a queue (RabbitMQ / SQS) better than a pub/sub (Kafka)?- What two patterns do you need to call an unreliable external API safely?- Explain CAP in one sentence and a one-line memory aid.- Why does every mutating API need an idempotency key?- How do you choose between hash and range sharding?- Why does read-replica lag matter for write-then-read?- Why use etcd or ZooKeeper for leader election?- How do you make observability a first-class design constraint, not an afterthought?

Python Fundamentals

• When is a generator better than a list comprehension?- Why reach for context managers over try/finally?- How do type hints + dataclasses work together?- Where is async/await the right tool?- When do you catch an exception, and when do you let it bubble?- How do you write a decorator that takes its own arguments?- When does collections outsmart a plain dict?- How do you mock external systems without breaking production code?- When f-strings win over .format / Template?- Where does PyPy or a JIT-ed build make sense?

SQL for AI Engineering

• How do you pick a row-per-group winner with RANK() vs DENSE_RANK()?- How does DISTINCT ON clean up duplicates without GROUP BY gymnastics?- How does ON CONFLICT (cols) DO UPDATE SCALE retries safely?- How do LAG/LEAD answer “what did this row look like yesterday”?- When do CTEs spark joy over nested subqueries?- How do you do top-K semantic search with pgvector?- Why JSONB @> containment is the right filter for AI metadata?- Why do a stratified train/test split in SQL instead of pandas?- Why compute features in SQL with shift-correct as of ts < cutoff?- How does EXPLAIN ANALYZE differ from a recall dashboard?

LLM & Prompt Engineering

• Does adding examples really change the model output?- System prompt vs user prompt — what is each for?- When do low temperature and high temperature help?- When does JSON mode beat regex post-processing?- Why use function calling for tool use instead of free text?- Does “let’s think step by step” actually help?- RAG or fine-tune?- When can LLMs replace humans in judging model output quality?- How do you budget an LLM’s context window?- What’s a reasonable per-request LLM cost and how do you control it?