[duclld1709]

Summary

Add a detailed interview preparation guide (INTERVIEW_QA.md) with 15 curated questions and answers that an AI/ML engineer might encounter when discussing this deepfake video detection project. The guide is structured to help candidates articulate technical decisions, debug findings, and architectural choices with concrete evidence from the codebase.

Key Content

15 Interview Questions covering:

• Project overview and problem statement (Q1)
• Critical bug fixes from v1→v2 refactoring:
  • LSTM batch_first=True bug and backward-compatible legacy mode (Q2-Q3)
  • Frame selection strategy and clip planning (Q4)
• Technical implementation details:
  • Face tracking with IoU-based continuity (Q5)
  • BGR color space preservation for train/serve consistency (Q6)
  • Grad-CAM explainability with memory optimization (Q7)
  • Inconclusive verdict calibration (Q8)
• Evaluation and validation:
  • Data leakage in v1 accuracy metrics (Q9)
  • Training improvements in v2 (Q10)
  • Augmentation strategy for temporal consistency (Q11)
  • Robustness evaluation methodology (Q12)
• Production considerations:
  • REST API design for concurrent inference (Q13)
  • Testing strategy without GPU/weights in CI (Q14)
• Self-awareness and roadmap (Q15)

Answer Structure:
Each question follows a consistent format:

• What the interviewer wants to verify
• Detailed answer with code references and file locations
• Follow-up questions to prepare for

Notable Features:

• Emphasizes the v1→v2 audit process as a strength, showing ability to find and fix bugs
• Grounds all answers in actual code (model.py, inference.py, training/train.py, tests/, etc.)
• Includes a quick reference table for common technical questions
• Honest about limitations (uncalibrated probabilities, missing cross-dataset benchmarks, legacy checkpoint constraints)
• Demonstrates understanding of ML-specific challenges: data leakage, train/serve skew, temporal consistency in video

https://claude.ai/code/session_01USUWxhjtWErys825QEgCLb