Interview practice plan
AI Architect
ZEISS · Bengaluru
Practice content written by LYDJ for interview preparation. These are common questions for this kind of role, not this employer's actual interview questions.
What this role asks for
Skills we found in the job description. The plan below is built around them.
Also named in the description, not covered below: Observability, Python, CI/CD. Review these on your own.
Sign in to see your gap areas first. We compare this role with the skills on your profile.
0 of 25 questions practised
Answer each one out loud, then tick it. Progress is saved in this browser only.
ETL & data pipelines
This role asks for ETL.
How do you design a pipeline that can be safely re-run for a past date?
What good looks likeHide hint
Idempotent writes via partition overwrite or merge on keys, deterministic inputs, no side effects outside the target, and a backfill procedure.
How do you detect and handle bad data arriving from an upstream source?
What good looks likeHide hint
Schema and data-quality checks (nulls, ranges, row counts, freshness), quarantining bad rows, alerting, and a data contract with the producer.
Batch or streaming: how do you decide for a new use case?
What good looks likeHide hint
How fresh the data must be, cost and operational complexity, late data handling, and whether micro-batches are good enough.
Agile
This role asks for JIRA. These Agile questions cover the ground it builds on.
How do you run sprint planning so the team commits to something realistic?
What good looks likeHide hint
A refined backlog, capacity based on real availability, historical velocity as a guide, a clear sprint goal, and room for unplanned work.
The team carries over stories every sprint. What do you do?
What good looks likeHide hint
Look at story size, work in progress, interruptions and dependencies, agree retro actions with owners, and track flow metrics such as cycle time.
Scrum or Kanban: when would you use each?
What good looks likeHide hint
Scrum for planned increments with a stable team; Kanban for continuous, interrupt-driven work such as support. Mention WIP limits.
Azure
This role asks for Azure.
How would you organise Azure management groups, subscriptions and resource groups for several teams?
What good looks likeHide hint
Management groups for policy inheritance, subscriptions as billing and isolation boundaries, resource groups per workload lifecycle, and consistent tagging.
When would you pick App Service, Azure Functions or AKS?
What good looks likeHide hint
Compare operational effort, scaling model, execution limits, cost profile and team skills, and give a workload that fits each.
How would you build a secure pipeline that deploys infrastructure and apps to Azure?
What good looks likeHide hint
Federated workload identity instead of stored secrets, Key Vault, environment approvals, infrastructure as code with Bicep or Terraform, and policy checks.
Audit
This role asks for Audit.
How would you plan an internal audit of the procure-to-pay process?
What good looks likeHide hint
Risk assessment, key controls (segregation of duties, approvals, three-way match, vendor master changes), sampling and testing of design and operation.
How do you choose a sample size, and what do you do when you find an exception?
What good looks likeHide hint
Base it on control frequency and risk; for an exception, understand the cause, consider extending the sample, assess impact and report it.
Explain SOX 404 and the role of IT general controls.
What good looks likeHide hint
Management's assessment of internal control over financial reporting; ITGCs (access, change management, operations) support the reliability of automated controls.
LLMs & GenAI
This role asks for Generative AI and RAG. These LLMs & GenAI questions cover the ground they share.
Explain retrieval-augmented generation. When does it beat fine-tuning?
What good looks likeHide hint
RAG grounds answers in fresh, citable documents without retraining; fine-tuning changes style or format behaviour. Many systems use both.
How would you evaluate an LLM-powered feature before launch?
What good looks likeHide hint
A golden dataset of real tasks, rubric-based and automated checks, human review of samples, a regression suite for prompt changes, and live feedback.
How do you reduce hallucinations in a question-answering system over company documents?
What good looks likeHide hint
Improve retrieval (chunking, hybrid search, reranking), instruct the model to answer only from sources with citations, allow it to abstain, and measure with evals.
SQL
This role asks for SQL.
Explain INNER, LEFT and FULL OUTER joins. When does a LEFT JOIN accidentally behave like an INNER JOIN?
What good looks likeHide hint
Filtering the right-hand table in WHERE removes the NULL rows; put that condition in the ON clause instead.
What is an index, how does a B-tree index speed up a query, and when does adding an index hurt?
What good looks likeHide hint
Sorted structure enabling seeks instead of scans; costs on writes and storage; low-selectivity columns and functions on columns can stop it being used.
A query that used to take 2 seconds now takes 2 minutes. How do you debug it?
What good looks likeHide hint
Read the execution plan, check stale statistics, missing or unused indexes, data growth and skew, parameter-sensitive plans, and locking.
Machine Learning
This is a data science / ML role. Expect questions on modelling choices, evaluation and production ML.
Explain the bias-variance trade-off with an example from a model you built.
What good looks likeHide hint
Underfitting versus overfitting, how you diagnosed it from training and validation curves, and what you changed (features, regularisation, data).
How do gradient-boosted trees work, and when would you prefer them over a neural network?
What good looks likeHide hint
Sequential trees fitting residuals; strong on tabular data with modest size, fast to train and easier to explain. Neural nets win on images, text and huge datasets.
How do you choose evaluation metrics for a ranking or recommendation problem?
What good looks likeHide hint
NDCG, MAP or recall@k depending on the use, aligned with a business metric, and awareness of the gap between offline and online results.
Behavioural
Every interview loop has a behavioural round. At mid level, expect questions on delivering independently, handling ambiguity and working through disagreement. Use the STAR structure.
Tell me about a mistake you made at work and what happened next.
What good looks likeHide hint
Own it without blaming others, explain how you contained the impact, and what you changed so it would not happen again.
Describe a time you had several urgent priorities at once. How did you decide what to do first?
What good looks likeHide hint
Explain your criteria (impact, deadlines, dependencies), how you communicated the trade-offs, and what you deliberately dropped or delayed.
Give an example of when you went beyond your role to fix a problem for your team or a customer.
What good looks likeHide hint
Pick something with real impact, explain why it mattered, and how you involved the owner rather than quietly working around them.
Tell me about a time you used data to change someone's mind.
What good looks likeHide hint
Explain what the data showed, how you made it understandable for that audience, and which decision changed as a result.
Company research brief: ZEISS
From the roles we track on LYDJ, refreshed daily from the official careers page.
110
open roles
41
added in 30 days
8
in Data & AI
Hiring most for
- Other39
- Engineering28
- Sales & Marketing20
- Data & AI8
Read before the interview
Questions to ask them
- ›There are several open Data & AI roles at ZEISS right now. How is the team growing, and where does this role fit?
- ›The company hires in Bengaluru, Delhi, Hyderabad. How do teams across locations work together day to day?
- ›How does the team use ETL and JIRA today, and what would you like to change?
- ›What does success look like in this role after six months?
- ›What is the biggest challenge the team is facing this year?
Preparation checklist
0/7 doneSaved in this browser only.