Context:

• Role: AI Engineer (IC2 level)
• Experience: ~4 Years
• Process: Phone Screen -> Virtual Onsite (5 Rounds)
• Outcome: Rejected (1-year cool-off)

Summary:
I had a very mixed loop. I received strong feedback on ML Theory and Product Vision for behavioural round during the interview, but I was rejected regareless without any feedback.

Round 0: Phone Screen (1 Hour)

• Format: Mix of ML Theory and light Coding.
• Question 1 (Theory): Explain the Attention Mechanism.
• Question 2 (Coding): Calculate the mean of an infinite data stream.
  • My Approach: Used the incremental formula to update the mean on the fly without storing all numbers (to handle infinite stream and prevent overflow).
  • Formula: ${\mu }_n={\mu }_{n-1}+\frac{x_n-{\mu }_{n-1}}{n}$
• Outcome: Passed.

Round 1: AI Coding (1 Hour)

• Format: AI-assisted IDE. Implement a LRU Cache data structure using AI tools and refine it.
• Question: Implement an LRU Cache. Later, add a validation method for keys.
• My Approach:
  • Used AI to generate the skeleton (Doubly Linked List + HashMap).
  • Prompted claude to implement a custom DoubleLinkedList class.
  • The interviewer kept saying add more features and ended up over engineering through just prompting.
• My taekaway: Not entirely sure what I was being evaluated for in this round.

Round 2: Data Structures & Algorithms (1 Hour)

• Question: First 30 minutes entirely on discussion of privious projects and experiences. Validate a Binary Search Tree (BST).
• My Approach:
  1. Solution 1: In-order traversal to check if the list is sorted. (Accepted, O(N)).
  2. Solution 2 (Optimization): Recursive approach to save space.

Round 3: ML Fundamentals & Concepts (1 Hour)

• Topics: Retrieval/Ranking theory, Imbalanced Data, Model Explainability.
• Outcome: Strong technical round. Deep theoretical understanding was a big plus. Even the interviewer complimented in theorital and analytical depth.

Round 4: ML System Design (1 Hour)

• Question: Design a "Second Pass Ranker" for the Feed serving billions of users. Constraint: No pre-existing embeddings.
• My Approach:
  • Designed a Two-Tower Architecture using Cosine Similarity.
  • Created seperate heads for the output layer (Like, Comment, Share).
• The Mistake:
  • Architecture Mismatch: Two-Tower is for Retrieval (Speed), not Ranking (Precision). For ranking 500 items, use DLRM or Wide & Deep.
  • Loss Function: Softmax implies mutual exclusivity. A user can Like AND Share. Use Multi-label Sigmoid heads.

Round 5: Hiring Manager (1 Hour)

• Focus: Product Vision, Behavioral, resume, and privious experiences.
• Outcome: HM complimented on my diverse industry experiences and projectes. Feedback was that I had "Strong Product Vision." Even suggested to chose any product facing teams during the team matching.

Iniitally the recruited said all rounds must go prefectly to be considered at LinkedIn level. So I am guess system design round killed my chance.