1. Q: Tell me about your Azure GenAI project.
A: I designed and implemented a scalable, agentic AI assistant using GPT-4 via Azure OpenAI. I
worked on LangChain agents, RAG integration with Azure AI Search and SQL DBs, built custom
tools and prompt templates, and deployed the system using Azure Functions for scale.
2. Q: How did you scale the AI system to 50,000 users?
A: I used Azure Functions for serverless deployment, auto-scaling compute, and added a 10%
bandwidth overhead buffer. Azure Blob Storage and SQL were tuned for latency, and DevOps
planning ensured robust cloud provisioning.
3. Q: What steps are involved in deploying an ML model to production?
A: Package model -> Build Docker image -> Push to registry -> Deploy to Cloud Run/GKE/Vertex AI
-> Expose endpoint -> Monitor usage and performance.
4. Q: What does your CI/CD pipeline look like?
A: Using Jenkins/GitLab: code pushed -> unit tests -> build + lint -> model validation -> Docker
image -> deployed to Cloud Run or GKE -> monitored with logs/alerts.
5. Q: How do you detect and mitigate data/model drift?
A: Track performance metrics and input distributions. Use statistical tests (e.g., PSI) and retrain
triggers if drift is detected. Alerting is done via monitoring tools like Prometheus or custom
dashboards.
6. Q: How do you extend LangChain functionality?
A: By creating custom tools for agent use, prompt routing logic, and multi-step reasoning chains. I've
built document loaders and integrated vector DB search with summarization pipelines.
7. Q: What is Retrieval-Augmented Generation (RAG)?
A: A GenAI system architecture where external documents (retrieved via semantic search) are
passed into the prompt context of the LLM to improve accuracy and reduce hallucinations.
8. Q: How did you fine-tune LLaMA 2 for finance?
A: Used LoRA (rank=64, alpha=16), 4-bit quantization, synthetic 48k Q/A dataset, trained on T4
GPUs via HuggingFace Transformers and PEFT in PyTorch.
9. Q: What do you monitor in production ML systems?
A: Input/output logs, latency, resource usage, accuracy over time, drift, and data pipeline failures.
Use dashboards, alerting tools, and audit trails.
10. Q: Which tools do you use for serving models?
A: Cloud Run (for serverless, lightweight APIs), GKE (for high-load models), Vertex AI Endpoints (for
managed ML deployment), Docker, and Flask/FastAPI for the inference service.
11. Q: How do you handle model versioning in MLOps?
A: I use Git for code and pipeline tracking, MLflow or custom registries for model versioning. Each
model artifact includes metadata (data version, hyperparams, metrics) to ensure reproducibility.
12. Q: How do you manage secrets in a deployment pipeline?
A: I use cloud-native solutions like GCP Secret Manager or Azure Key Vault. Secrets are injected at
runtime through environment variables, never hardcoded.
13. Q: What is a canary deployment in MLOps?
A: It's a deployment strategy where a new model version is exposed to a small percentage of users
to monitor for regressions before full rollout.
14. Q: How would you monitor a GenAI application like a chatbot?
A: Track latency, token usage, user feedback, prompt-token ratio, hallucination rate, and failures.
Tools: Azure Monitor, custom logging with alerts.
15. Q: What is a feature store and why is it important?
A: A centralized system to store, serve, and reuse features across ML models. Ensures consistency
between training and inference.
16. Q: How have you used Apache Airflow in your pipelines?
A: Used GCP Composer (Airflow) for orchestrating ETL jobs, model retraining workflows, and
scheduling data validation tasks.
17. Q: Why is Docker important in MLOps?
A: Docker ensures environment consistency across dev, test, and prod. Helps package
dependencies, models, and apps into reproducible containers.
18. Q: How do you handle explainability in ML models?
A: For critical models (e.g. medical/finance), I use SHAP, LIME, or attention visualization to
understand feature influence and gain user trust.
19. Q: How do you optimize ML workloads on cloud?
A: Use spot/preemptible instances, autoscaling, quantization (LLMs), and GPU-aware scheduling.
Minimize idle compute via triggers or Functions.
20. Q: How do you adapt your MLOps pipeline between Azure and GCP?
A: Abstract components using modular code. For example, switch between Azure ML and Vertex AI
by adjusting SDKs and auth, while keeping pipelines generic.