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

What the job involves

Computer Vision Engineers at AI and deeptech companies like Sarvam AI build and deploy models that extract meaning from images and videos, often for Indian language and document understanding tasks. Day to day, they work on data pipelines, train and fine-tune vision models, optimize inference for edge or cloud, and collaborate with product teams to ship features. The role blends deep learning research with production engineering, requiring strong fundamentals in image processing, deep learning, and software development.

Role and Technical Knowledge

1. 1

   Explain the difference between object detection and image segmentation. When would you use each?

   Detection gives bounding boxes and class labels; segmentation provides pixel-level masks. Use detection for counting or coarse localization, segmentation for precise boundaries like medical imaging or background removal.

2. 2

   How does a convolutional layer work? What are receptive fields and why do they matter?

   Convolution applies filters to local patches, sharing weights. Receptive field is the region of input affecting a neuron; deeper layers have larger receptive fields, enabling capture of global context.

3. 3

   What is the role of attention in Vision Transformers (ViTs) compared to CNNs?

   ViTs split images into patches and use self-attention to model global relationships from the start, while CNNs build locality inductively. ViTs often need more data but excel at capturing long-range dependencies.

4. 4

   Describe how you would handle class imbalance in an image classification dataset.

   Mention techniques like oversampling, undersampling, class weighting, focal loss, and data augmentation. Also consider evaluation metrics like F1-score or AUC instead of accuracy.

Practical Scenarios and Problem Solving

1. 5

   You need to deploy a real-time object detection model on a mobile device with limited compute. What steps would you take?

   Discuss model selection (e.g., YOLO, MobileNet), quantization, pruning, knowledge distillation, and using frameworks like TensorFlow Lite or ONNX Runtime. Mention trade-offs between accuracy and latency.

2. 6

   Your model performs well on validation but poorly on real-world images from Indian streets. How would you debug and improve it?

   Check for domain shift, collect diverse real-world data, analyze failure cases, augment with relevant transformations, and consider fine-tuning on target domain. Also verify data preprocessing and label quality.

3. 7

   How would you build a pipeline to extract text from scanned Indian documents in multiple languages?

   Outline steps: image preprocessing (denoising, deskewing), text detection (e.g., EAST, DBNet), recognition (CRNN, TrOCR), and language-specific OCR models. Mention handling of Devanagari and other scripts.

4. 8

   You have a dataset of 10,000 images but only 500 are labeled. How would you approach training a model?

   Suggest semi-supervised learning, self-supervised pretraining, data augmentation, active learning, or transfer learning from a pretrained model. Emphasize starting with a pretrained backbone and fine-tuning on labeled data.

Behavioural and HR Questions

1. 9

   Tell me about a computer vision project you worked on. What was your role and what challenges did you face?

   Use STAR format: describe the problem, your approach, technical hurdles (e.g., data scarcity, model convergence), and the outcome. Highlight learnings and teamwork.

2. 10

   Why do you want to work in computer vision at a company like Sarvam AI?

   Show passion for AI and deeptech, mention interest in Indian language and document understanding, and align with the company's mission. Avoid generic answers; be specific about the sector.

3. 11

   Describe a time you had to learn a new technology quickly to solve a problem.

   Pick a concrete example, such as learning a new framework or algorithm under deadline. Explain your learning process, how you applied it, and the result.

4. 12

   How do you handle feedback or criticism on your work?

   Demonstrate openness, give an example where you improved based on feedback, and emphasize a growth mindset and collaboration.

Coding and Implementation

1. 13

   Write a function to perform image augmentation (e.g., random rotation and flip) using OpenCV or PIL.

   Show clean code with random angle selection, rotation, and flipping. Mention preserving image dimensions and handling bounding boxes if applicable.

2. 14

   Implement a simple convolutional layer forward pass in NumPy.

   Demonstrate understanding of sliding window, padding, stride, and output shape calculation. Use loops or im2col for efficiency.

3. 15

   Given a list of bounding boxes, write code to compute Intersection over Union (IoU) between two boxes.

   Compute intersection area, union area, and divide. Handle edge cases like no overlap. Show vectorized implementation if possible.

4. 16

   How would you load and preprocess a dataset of images for training in PyTorch? Write a snippet.

   Use torchvision.datasets.ImageFolder, DataLoader, and transforms (Resize, ToTensor, Normalize). Mention batching, shuffling, and parallel loading with num_workers.