AI Engineering Project Idea and GenAI Interview Questions with Answers

Put every lesson together in one portfolio project that companies understand immediately.

1. The project: Campus Help-Desk Assistant

Users: students and parents. Goal: answer questions about rules, fees, exams and attendance correctly, from official documents, and look up a student's own data safely.

Part	Lesson	What you build
Documents → chunks → index	6, 7	handbook sections with section numbers; hybrid search
Grounded answers with citations	3, 7	prompt template v1 in a file; "not in the handbook" path
Student data via tools	8	`get_attendance(roll)` — only for the logged-in student
Structured ticket creation	4	out-of-scope questions become tickets with a schema
Chat with history and streaming	5	last 6 turns + summary; streaming in the web UI
Guardrails	11	redaction, injection screening, no write tools without approval
Cost and limits	10	cached system prompt, per-user daily limit, usage logs
Evaluation	9	50 real questions with expected sections/labels; run on every change

Stack suggestion: Python (FastAPI) or Spring Boot backend, PostgreSQL + pgvector, a simple React or plain HTML front-end, the Claude API, and GitHub Actions running your eval script.

README must include: the problem, architecture diagram, how to run it (including with the free local stand-in API), evaluation results (e.g. "retrieval hit rate @ 3: 47/50"), known limits, and cost per 1,000 questions.

2. A tiny but complete core

The heart of the assistant — routing a question to the right path — fits in a few lines and is easy to test:

def route(question: str, logged_in_roll: int | None) -> str:
    q = question.lower()
    if any(w in q for w in ("my attendance", "my marks", "my result")):
        return "tool:get_attendance" if logged_in_roll else "ask_login"
    if any(w in q for w in ("fee", "attendance", "hall ticket", "exam", "re-evaluation")):
        return "rag:handbook"
    return "ticket:office"

cases = [
    ("What is my attendance in DSA?", 104),
    ("What is my attendance in DSA?", None),
    ("What is the late fee?", None),
    ("Can the canteen open on Sunday?", 104),
]
for question, roll in cases:
    print(f"{question:32s} roll={roll!s:5s} -> {route(question, roll)}")

What is my attendance in DSA?    roll=104   -> tool:get_attendance
What is my attendance in DSA?    roll=None  -> ask_login
What is the late fee?            roll=None  -> rag:handbook
Can the canteen open on Sunday?  roll=104   -> ticket:office

A rule-based router like this is cheap and predictable; later you can replace it with a model call (classification with structured output) — and your eval set tells you whether that's actually better.

3. Interview questions — short answers

What is an LLM? A neural network trained to predict the next token; instruction-tuned to follow requests.
What is a token / context window? Word pieces; the maximum input+output the model can handle in one call.
Why do LLMs hallucinate? How do you reduce it? They generate plausible text, not looked-up facts. Ground with RAG, use tools, allow "I don't know", cite sources, evaluate.
Explain RAG. Chunk and index documents, retrieve relevant chunks for each question, put them in the prompt, answer with citations.
Chunk size trade-off? Too big: unfocused and costly; too small: loses context. Use sections/paragraphs with slight overlap; measure.
What are embeddings? Vectors representing meaning; similar texts are close (cosine similarity).
Vector database? Stores vectors with metadata and does fast nearest-neighbour search (pgvector, Chroma…).
Keyword vs semantic search? Exact words (good for codes/names) vs meaning (good for paraphrases); hybrid combines both.
What is tool use / function calling? The model requests a function with arguments; your code runs it and returns the result.
Agent vs workflow? The model chooses the steps vs your code fixes the steps; prefer workflows unless steps can't be known.
How do you get reliable JSON? Structured outputs with a JSON schema / Pydantic, plus your own validation.
How do you evaluate an LLM app? A versioned test set of real cases; exact/rule/reference scoring, LLM-as-judge validated against humans; compare versions.
What is prompt injection? Defences? Untrusted text that contains instructions. Tag data, least privilege, output validation, human approval.
How do you cut cost? Prompt caching, fewer input tokens, shorter outputs, right model per task, batching.
How do you reduce latency? Streaming, shorter prompts, caching, parallel calls, lighter models for simple tasks.
Fine-tuning vs RAG vs prompting? Prompting first; RAG for knowledge; fine-tuning (where available) for style or narrow formats — rarely the first step.
What would you log? Prompt version, model, tokens, cost, latency, stop reason, sources, tool calls, feedback — minimal personal data.
How do you handle personal data? Minimise and redact, follow the DPDP Act and provider policies, limit retention.
Why keep humans in the loop? High-stakes or irreversible decisions; accountability; catching model errors.
Describe your project. Problem → architecture → evals with numbers → a failure you found and fixed → cost and limits.

Common mistakes

A demo with no evaluation numbers.
Keys in the repo; no rate limits on a public demo.
Claiming the model "understands" or "never makes mistakes".

Quick recap

Capstone: RAG + tools + structured tickets + chat + guardrails + cost controls + evals.
Start rule-based where you can; prove improvements with your eval set.
Interviews reward numbers, trade-offs and honest limits.

Capstone — Build a Campus Help-Desk Assistant, Plus AI Engineering Interview Questions