Murf AI Computer Vision Engineer Interview: Rounds, Questions and Preparation

What the job involves

Computer Vision Engineers in AI and deeptech build and deploy models that extract meaning from images and video, often for products like voice avatars, video dubbing, or visual quality checks. Day to day, you will preprocess data, train and fine-tune models such as CNNs, transformers, or GANs, and optimise them for real-time inference on cloud or edge devices. You will also collaborate with product and ML teams to turn research prototypes into reliable features.

Role and technical knowledge

1. 1

   Explain how a convolutional layer works and why it is useful for images.

   Cover local receptive fields, weight sharing, and translation invariance. Mention how filters learn edges and textures in early layers.

2. 2

   What is the difference between object detection and image segmentation?

   Detection outputs bounding boxes and class labels; segmentation assigns a class or instance mask per pixel. Give examples like YOLO for detection and Mask R-CNN for segmentation.

3. 3

   How would you handle a dataset with very few labelled images for a CV task?

   Discuss transfer learning, data augmentation, semi-supervised learning, and synthetic data. Mention starting with a pretrained backbone like ResNet or ViT.

4. 4

   What metrics would you use to evaluate a face or lip-sync model for a video dubbing product?

   Mention pixel-level metrics like L1 or SSIM, perceptual metrics like LPIPS, and task-specific metrics such as lip-sync confidence or landmark error.

Practical scenarios

1. 5

   Your model performs well on validation but poorly in production on user-uploaded videos. How do you debug this?

   Check for distribution shift, compression artifacts, lighting, and frame rate differences. Suggest building a monitoring pipeline and retraining with production-like data.

2. 6

   You need to run a CV model on a mobile device with limited compute. What steps do you take?

   Discuss model pruning, quantization, knowledge distillation, and converting to ONNX or TFLite. Mention testing latency and accuracy trade-offs.

3. 7

   How would you build a pipeline to detect and blur background in a real-time video call?

   Outline person segmentation with a lightweight model like MobileNetV3 or MediaPipe, temporal smoothing across frames, and efficient post-processing.

4. 8

   A teammate claims their new augmentation improves accuracy, but your test shows no gain. How do you resolve this?

   Suggest running a controlled experiment with the same seed and splits, checking statistical significance, and reviewing whether the augmentation matches the test distribution.

Behavioural and project questions

1. 9

   Tell me about a computer vision project you built. What was the hardest part?

   Use the STAR method. Focus on a specific technical challenge like data cleaning, class imbalance, or latency, and explain what you learned.

2. 10

   Describe a time you had to learn a new CV tool or paper quickly for a deadline.

   Show a concrete example of reading a paper, reproducing results, and applying it. Highlight time management and initiative.

3. 11

   How do you explain a complex CV model to a non-technical product manager?

   Use analogies and focus on inputs, outputs, and business impact. Avoid jargon and confirm understanding.

4. 12

   Why do you want to work in AI and deeptech, and why this role?

   Connect your interest to real products like voice avatars or video dubbing. Mention your skills and desire to ship models that users interact with.

HR and company fit

1. 13

   What do you know about our company and the CV problems we might solve?

   Research the company's products, such as voice synthesis or video localisation, and link them to CV tasks like lip-sync, face reenactment, or quality control.

2. 14

   Where do you see yourself in three years as a computer vision engineer?

   Show a growth path from implementing models to leading a small CV module or owning a production pipeline. Align with the company's scale.

3. 15

   How do you handle feedback when your model or code is criticised?

   Emphasise listening, asking clarifying questions, and iterating quickly. Give an example of improving after review.

4. 16

   Are you comfortable working from an office or in a hybrid setup, and do you have any location constraints?

   Be honest and flexible. Mention willingness to relocate if the role requires it and you are open to it.