Synthesia India NLP Engineer Interview: Rounds, Questions and Preparation

What the job involves

As an NLP Engineer in AI and deeptech, you build and fine-tune language models for tasks like text generation, summarisation, and conversational AI. Day to day involves data preprocessing, model training and evaluation, and deploying NLP pipelines into production. You work closely with ML engineers, product managers, and designers to ship language features that users interact with.

Role and Technical Knowledge

1. 1

   Explain the self-attention mechanism in transformers and why it is useful for NLP tasks.

   Cover how queries, keys, and values compute weighted sums, and mention parallelisation and long-range dependency capture.

2. 2

   What is the difference between BERT and GPT in terms of architecture and training objective?

   Contrast masked language modelling vs causal language modelling, and discuss bidirectional vs unidirectional context.

3. 3

   How would you handle out-of-vocabulary words when using a pre-trained tokenizer?

   Mention subword tokenization like BPE or WordPiece, and the use of unknown tokens or byte-level fallbacks.

4. 4

   What evaluation metrics would you use for a text summarisation model, and what are their limitations?

   Discuss ROUGE, BLEU, and human evaluation, and note that n-gram overlap may not capture factual correctness or fluency.

Practical Scenarios and Problem Solving

1. 5

   You have a dataset of customer support tickets in English and Hindi. How would you build a model to classify them into categories?

   Talk about multilingual models like mBERT or XLM-R, data preprocessing, handling code-mixing, and evaluation per language.

2. 6

   Your fine-tuned LLM is generating repetitive and factually incorrect responses. How would you debug and improve it?

   Suggest checking training data quality, adjusting decoding parameters like temperature and top-p, and using reinforcement learning from human feedback or retrieval augmentation.

3. 7

   How would you deploy an NLP model to serve real-time predictions with low latency?

   Mention model quantization, ONNX runtime, batching, caching, and using a serving framework like TorchServe or Triton.

4. 8

   You need to build a named entity recognition system for Indian legal documents with limited labelled data. What approach would you take?

   Propose transfer learning from a pre-trained model, active learning, data augmentation, and possibly rule-based post-processing.

Behavioural and HR Questions

1. 9

   Tell me about a time you worked on a team project and faced a disagreement. How did you resolve it?

   Use the STAR method, focus on communication and compromise, and highlight the positive outcome for the project.

2. 10

   Why do you want to work as an NLP Engineer in the AI and deeptech sector, and why our company?

   Show passion for language AI, mention specific products or research from the company, and align your skills with their mission.

3. 11

   Describe a challenging technical problem you solved. What was your approach?

   Pick a project where you had to learn something new, break down the problem, and iterate. Emphasise persistence and learning.

4. 12

   Where do you see yourself in three years, and how does this role fit into your career goals?

   Express a desire to grow into a senior NLP engineer or tech lead, contributing to impactful language technologies.

Coding and Implementation

1. 13

   Write a Python function to compute the cosine similarity between two sentence embeddings.

   Use numpy or PyTorch, handle zero vectors, and explain the formula dot product divided by product of norms.

2. 14

   Given a list of strings, write code to tokenize them using a Hugging Face tokenizer and pad the sequences.

   Show familiarity with AutoTokenizer, batch encoding, padding and truncation arguments, and return tensors.

3. 15

   Implement a simple attention layer in PyTorch.

   Define linear layers for query, key, value, compute scaled dot-product attention, and apply softmax. Discuss masking.

4. 16

   How would you fine-tune a pre-trained model like DistilBERT for sentiment analysis using PyTorch?

   Outline loading the model, adding a classification head, setting up DataLoader, training loop with optimizer and loss, and evaluation.