How would you build an AI assistant that works well in several Indian languages?
Prompting, RAG or fine-tuning · Pro 1
Short answer
Start with a test set of real user messages in each language and script, including code-mixed and Romanised text, checked by native speakers. Compare a general model, translate-to-English-and-back, and an Indian-language model on quality, tokens and cost per language. Retrieve local facts and names in each language, and monitor every metric per language after launch.
Deeper answer
Quality differs a lot by language: models are often decent in Hindi and weaker in lower-resource languages like Odia or Assamese, because those are rare in training data. Translation can help weak languages but loses nuance such as respect forms and can garble names.
Indian groups such as AI4Bharat publish open translation, speech and evaluation resources, and the government's Bhashini platform offers language services. For voice, I would test speech recognition with real recordings and accents. I would keep glossaries of product and scheme names, reply in the user's language, and re-test after every model change.
What a strong answer covers
- Real test set per language and script, native speakers
- Compares approaches
- Considers tokens and cost
- Retrieves local facts and glossaries
- Monitors per language
Answer it out loud in the mock interview about 1.5 min
Follow-ups to expect
- Which metrics would you report per language?
- How would you handle Hinglish?
Common mistakes
- Testing only in English or formal Hindi
- Treating all Indian languages as one
Learn it
AI in Indian languages: tokens, quality gaps and what to use
- How would you build a dataset to fine-tune an LLM for a business task?
- Your fine-tuned model is no better than the base model. How do you debug it?
- Models now have very large context windows. Why not put all the documents in the prompt instead of using RAG?
- You have three LoRA adapters for three tasks on the same base model. Would you merge them, serve them separately, or train one multi-task model?
- How do you decide which model size to use for a feature?
- How would you choose a base model to continue training or fine-tune for a specialised domain or language?
- What is prompt caching, and when does it help?
- When would you use prompting, RAG or fine-tuning?
- What is model distillation, and when would you use it?
- When is fine-tuning worth the effort?
- A reasoning model is more accurate on your task but five times the cost and much slower. How do you decide whether to use it?
- When would you use synthetic data to train or test a model, and what are the risks?