Based on real interview feedback across Persistent, Infosys Topaz, Genpact, Accenture, Globant and similar enterprise GenAI teams. The pattern is clear: Python first, project depth next, then production RAG, cloud, microservices, latency, evaluation, security and guardrails.
Candidates are not rejected only because they do not know RAG or LLMs. Many candidates know the words. The real problem is that they cannot explain implementation depth, production tradeoffs, latency, cloud services, evaluation, security and guardrails in a structured way.
The real interview journey moves from coding to project explanation, then to RAG, multimodal PDFs, production engineering and enterprise-grade safety.
The interviews follow a repeated structure across companies. The names may change, but the pattern remains similar.
Python + Project Screening
Production GenAI Deep Dive
Reusable Enterprise GenAI Solutions
The discussion usually starts simple and then becomes production-oriented.
Basic programs using strings, lists, dictionaries, loops, functions, file handling, JSON and APIs.
Explain business problem, architecture, data flow, model choice, APIs, database, deployment and outcomes.
Chunking, embeddings, vector DB, retrieval, reranking, prompt construction, citations and hallucination control.
Handle text, tables, images, stamps, charts, scanned pages and multi-column layouts.
Latency, cost, caching, async calls, microservices, FastAPI, Docker, CI/CD and cloud services.
Faithfulness, answer relevance, context precision, security, PII masking, prompt injection defense and access control.
L1 is usually lighter. The interviewer may ask you to open an online editor and solve simple Python programs. Even pseudocode may be acceptable if logic is clear.
| Topic | Why it matters |
|---|---|
| String and list problems | Most L1 rounds start with simple coding checks. |
| Dictionary and set usage | Useful for frequency count, deduplication and quick lookup. |
| Functions and clean logic | Shows structured problem-solving and reusable code thinking. |
| File and JSON handling | Common in document, API and GenAI data pipelines. |
| Exception handling | Important when external LLM APIs, files or network calls fail. |
| OOP basics | Helps explain services, classes, interfaces and maintainable architecture. |
| Async, threads and processes | Important for latency, API calls and performance design. |
Interviewers do not ask only "What is RAG?" They ask how RAG was implemented in your project.
| RAG Area | What the candidate must explain |
|---|---|
| Data Source | PDFs, websites, internal documents, databases, images, invoices or enterprise knowledge bases. |
| Ingestion | How documents are uploaded, parsed, cleaned, versioned and stored. |
| Chunking | Why fixed, recursive, semantic, section-based, parent-child or table-aware chunking was selected. |
| Embeddings | Which embedding model was used, why it was selected and how embedding quality was measured. |
| Vector Database | Pinecone, Chroma, FAISS, pgvector, OpenSearch or Azure AI Search and why it fits the use case. |
| Retrieval | Top-k search, hybrid search, metadata filters, access-aware retrieval and query rewriting. |
| Reranking | When reranking improves relevance and when it increases latency unnecessarily. |
| Prompting | Grounded prompts, source citations, fallback behavior and refusal when context is insufficient. |
| Evaluation | Faithfulness, answer relevance, context precision, context recall, hallucination rate and latency. |
| Guardrails | PII masking, prompt injection defense, role-based access, policy filters and audit logs. |
Enterprise PDFs contain text, tables, charts, stamps, scanned pages and images. A plain-text RAG answer is not enough.
I will split the PDF into chunks, create embeddings and store them in a vector database.
I will classify PDF content into text, tables, images and scanned pages. I will extract text with layout awareness, extract tables structurally, use OCR or vision models for images, create metadata-rich chunks and retrieve relevant text, table and image summaries together.
A weak answer says "I used 500 tokens." A strong answer selects chunking based on document type and retrieval goal.
| Chunking Strategy | Best Used For | What can go wrong |
|---|---|---|
| Fixed-size chunking | Simple plain text | May split important meaning across chunks. |
| Recursive chunking | General documents | May still ignore document layout. |
| Semantic chunking | Meaning-based sections | Can be slower and needs careful tuning. |
| Section-based chunking | Policies, manuals and reports | Weak headings can create poor chunks. |
| Clause-based chunking | Contracts and legal documents | Bad splitting can change legal meaning. |
| Table-aware chunking | Invoices, financial reports and tabular PDFs | Rows and columns may lose context if flattened badly. |
| Parent-child chunking | Long documents where context matters | Needs more storage and retrieval logic. |
| Sliding-window chunking | Documents requiring overlap | Increases token cost and duplicate retrieval. |
L2 interviewers usually have higher experience and want practical answers. They ask about latency, microservices, FastAPI, cloud services and implementation tradeoffs.
Use FastAPI when the GenAI application is API-first, needs async support, low-latency endpoints, automatic OpenAPI documentation and easy integration with Python ML libraries. Use Django when you need a full web framework with ORM, admin panel, authentication and server-rendered pages.
Production GenAI systems must be measured, not only built.
| Metric | Meaning |
|---|---|
| Average latency | Normal response time across requests. |
| P95 latency | 95% of users get a response within this time. |
| Time to first token | How quickly the streamed answer begins. |
| Retrieval latency | Time taken by vector or hybrid search. |
| LLM latency | Time taken by the language model to generate. |
| End-to-end latency | Total time from user question to final response. |
| Cost per query | Total retrieval, model and infrastructure cost per answer. |
Do not say only "I used cloud." Explain the service layer and why it was selected.
S3, Bedrock, OpenSearch, Aurora pgvector, ECS/EKS, Lambda, API Gateway, IAM, KMS, Secrets Manager, CloudWatch
Azure OpenAI, Azure AI Search, Blob Storage, App Service, Functions, AKS, Entra ID, Key Vault, Azure Monitor, App Insights
Vertex AI, Gemini, Cloud Storage, Cloud Run, Cloud SQL, Firestore, Vertex AI Search, IAM, Secret Manager, Cloud Monitoring
Globant-style rounds can ask Python programming methodologies, threads, processes and async programming. Candidates must know when to use which.
| Approach | Best For | Interview Line |
|---|---|---|
| Multithreading | I/O-bound work like API calls, database calls and file reads. | Use threads when tasks are waiting on network or disk. |
| Multiprocessing | CPU-bound work like heavy computation, image processing or large preprocessing. | Use processes when CPU work must run in parallel. |
| Async programming | High-concurrency I/O like many LLM or retrieval calls. | Use async when the service handles many waiting operations efficiently. |
When asked "How many microservices did your application have?", explain responsibilities and scaling needs.
| Microservice | Responsibility |
|---|---|
| User Service | Login, roles, profiles and authentication. |
| Document Ingestion Service | Upload, parse, clean and version documents. |
| Embedding Service | Create embeddings and manage embedding model versioning. |
| Retrieval Service | Search vector DB, apply filters and return relevant context. |
| LLM Service | Call the LLM and generate grounded responses. |
| Guardrail Service | Check safety, policy, PII and prompt injection risk. |
| Evaluation Service | Score faithfulness, relevance, retrieval quality and hallucination risk. |
| Feedback & Monitoring Service | Capture ratings, logs, latency, cost, errors and user feedback. |
We separated ingestion, retrieval, generation, guardrails and evaluation because each service has different scaling needs. Ingestion is batch-heavy, retrieval needs low latency, generation depends on LLM APIs, guardrails need strict policy checks and evaluation can run asynchronously.
This is the most important signal from the interview feedback. Candidates must prepare this deeply.
| Area | What it checks |
|---|---|
| Faithfulness | Whether the answer is supported by retrieved context. |
| Answer relevance | Whether the answer directly solves the user question. |
| Context precision | Whether retrieved chunks are useful and not noisy. |
| Context recall | Whether retrieval found all required information. |
| Hallucination rate | How often the model invents unsupported facts. |
| Citation accuracy | Whether citations map to the correct source. |
| Authentication | Whether users are properly identified. |
| Authorization | Whether users can access only allowed documents. |
| PII masking | Whether sensitive personal data is hidden or protected. |
| Prompt injection defense | Whether malicious instructions inside documents are ignored. |
| Audit logs | Whether user questions, retrieved docs and system decisions are traceable. |
| Encryption | Whether data is protected at rest and in transit. |
Use this structure for RAG, cloud, architecture, guardrails, latency and project explanation answers.
What business problem was being solved?
What components were used and why?
How was it built step by step?
Why this design over alternatives?
How were accuracy, latency, safety and cost measured?
How were hallucination, data leakage and failures handled?
What was tracked after production release?
How did the solution save cost, reduce time or improve quality?
Candidates should prepare in this priority order before appearing for enterprise GenAI interviews.
Prepare 30-40 small coding problems covering strings, lists, dictionaries, files, JSON, APIs, exception handling, OOP, async, threads and processes.
Prepare one strong GenAI project with architecture diagram, data flow, APIs, database, RAG design, evaluation, guardrails, deployment and monitoring.
Revise chunking, embeddings, vector DB, hybrid search, reranking, metadata filtering, multimodal PDFs, query rewriting, citations and hallucination control.
Prepare FastAPI, microservices, Docker, cloud services, CI/CD basics, logging, monitoring, latency optimization, cost optimization and scaling.
Prepare RAGAS-style evaluation, faithfulness, context precision, answer relevance, PII masking, access control and prompt injection defense.
The candidate prepares definitions and tool names only.
The candidate prepares implementation and production-level answers.
The Talent Grid helps candidates convert scattered GenAI knowledge into structured, role-ready interview performance with Python coding, project deep-dive, production RAG, cloud, microservices, evaluation and guardrails.
