Background:

• Education: Final year B.Tech CSE, Indian Institute of Information Technology (IIIT) Bhagalpur
• Role: SAP STAR Program
• Prior Experience: Training and Placement Coordinator, Active Hackathon Participant.
• Core Expertise: Full Stack (MERN), AI, Machine Learning, RAG pipelines, Vector Databases, and System Design.

I recently went through the SAP STAR Program interview process. Although I was rejected after the final HR round, the technical deep-dives were incredibly rigorous and rewarding. It was a masterclass in architectural reasoning rather than rote memorization. Here is a comprehensive breakdown of my experience to help future candidates.

Round 1: Technical Interview (1 Hour | 2 Interviewers)

The interview was highly conversational. At the very start, since my resume spanned multiple domains, the interviewers asked me to pick my top two strongest areas to drive the conversation. I chose AI and Full Stack.

1. Projects & Architecture Discussion

The interview kicked off by exploring my projects, with a heavy focus on my LLM + RAG project, AnswerDeck.

• Authentication Flow: We discussed modern authentication flows. They asked why I implemented a custom JWT-based authentication system rather than relying on session-based alternatives.
• Follow-up Questions: What is JWT? Why use it? How does the authentication flow work under the hood? What is the fundamental difference between JWT and Sessions?

2. AI Architecture & Generative AI Deep Dive

From there, the interview pivoted entirely into modern AI architecture, specifically testing practical product intuition.

• GraphRAG & Business Intuition: I was given a situational business use-case designed to test Knowledge Graphs vs. RAG. I pitched GraphRAG. I was completely honest that I had only read about the architecture and had not implemented it yet. Instead of dismissing it, the interviewer used it to test my raw intuition on how to represent entity relationships and chunk data in a graph environment.
• Chunking Strategies (Live Pseudocode): I was asked to write pseudo-code for a basic text chunking strategy.
  • Follow-up: Modify the code to implement a Sliding Window Chunking strategy. (I successfully implemented this on the fly by applying fundamental DSA sliding window concepts).
  • Follow-up: "If not sliding window or fixed chunk, what chunking strategy is commercially used by major RAG implementations?" We discussed recursive, semantic, and document-aware chunking with overlap. We also explored how chunk overlap preserves context differently than relationships in a Knowledge Graph.
• RAG Pipeline & Embeddings: They asked me to explain the complete RAG pipeline (Document -> Chunking -> Embedding -> Vector Search -> Top-k Retrieval -> LLM -> Response). We discussed embedding generation, cosine similarity, and semantic reranking.
• Context Window Optimization: A major discussion on how to minimize context window token usage. We discussed prompt optimization, conversation summarization, context compression, and memory management.
• LLM Under the Hood: Is an LLM stateful or stateless? If stateless, how do applications like ChatGPT remember conversation history? We discussed the differences between the stateless model itself, the application layer, prompt history, and external memory.
• AI Hallucination: Is hallucination inversely proportional to prompt quality? We discussed how prompting helps, but hallucination fundamentally depends on model limitations, grounding, and validation.

3. Machine Learning & Fundamentals

• Logistic vs. Linear Regression: We discussed the differences, continuous vs. binary output, and real-world classification examples. I was asked to explain the activation function used in Logistic Regression and provided the Sigmoid formula:
  $\sigma (x)=\frac{1}{1+e^{-x}}$
• Frontend vs. Backend Performance: The interviewer stated, "Good backend health directly impacts frontend performance." I agreed, discussing how backend latency, API efficiency, database optimization, and caching play a role, but also balanced the argument by highlighting frontend optimizations like lazy loading, rendering, and bundle size reduction.

4. Logic Puzzles & Problem Solving

The interviewers tested logical reasoning through classical puzzles.

• Puzzle 1: River raft with 4 people and minimum time. (The interviewer joked that past candidates probably shared this online, so we skipped it after a laugh).
• Puzzle 2 (Poisoned Wine Problem): 1000 bottles, 1 poisoned, 10 rats, 1 day to find the bottle.
  • Concept (Binary Encoding): Since $2^{10}=1024$, 10 rats are sufficient to represent every bottle in binary combinations. I recognized the concept because I had practiced it, and I was completely honest about it. They deeply appreciated the transparency.
• Puzzle 3 (Two Guards Problem): Two doors, two guards (one truthful, one liar). Find the diamond with one question.
  • Concept: Ask, "If I asked the other guard which door has the diamond, what would he say?" and choose the opposite. (I struggled with this one and could not arrive at the final answer).

5. Career & The "Jack of All Trades" Reality

Towards the end, we had a candid conversation about focusing on one domain vs. exploring many. I shared my perspective: in today’s volatile market, as a fresher, having a broad grasp across Full Stack and AI is often a necessary survival strategy to secure initial employment while still feeding my enthusiasm for building end-to-end systems. They completely agreed and appreciated the reality check.

Round 2: HR & Behavioral (1 Hour | 3 Interviewers)

After a short break, I had the HR round. I made sure to stay calm and deliver a strong, confident introduction. This wasn't a generic HR round; it heavily tested ethics, ownership, and conflict resolution.

• Questions Asked: Why SAP? How does SAP align with your career goals?
• Situational Scenarios: Walk us through your team projects. Drawing from my experience as a Training and Placement Coordinator, I navigated a scenario about teammate drop-offs during a hackathon (where a teammate stops contributing but knows the evaluators personally). We discussed conflict resolution, maintaining integrity, and why one should never misuse personal relationships with evaluators.
• General Discussion: The HR head seemed very impressed with my introduction and clarity of thought regarding professional ethics, teamwork, and a learning mindset.

The Reality & Key Takeaways

Getting rejected at the finish line three times in a row is exhausting. It is difficult not to feel the sting. However, the experience was a phenomenal confidence booster and a masterclass in how industry ML and systems actually work.

For Future Candidates:

1. Do not memorize definitions: Interviewers will constantly ask "Why?", "What happens internally?", "What are the tradeoffs?", and "What is the implementation cost?"
2. DSA isn't just for LeetCode: Seeing how core Data Structures and Algorithms directly apply to modern GenAI pipelines (like sliding window chunking) was a huge realization.
3. Honesty wins: Admitting I hadn't implemented GraphRAG, and being truthful about seeing puzzles on AI tools, built immense trust.
4. Think from production scale: Understand RAG beyond basic tutorials. Think about commercial viability, backend optimization, and state management.

I am taking a short reset and going back to the grind from scratch—revising DSA, shipping code, and deep-diving into System Design. To everyone currently navigating this placement season: let's keep pushing. The right door will open.