Company Interview Questions · AI Labs & Startups

Krutrim & Sarvam AI: Indian AI Labs

❓ 6 questions🏢 Krutrim & Sarvam AI🆕 Recent reports

Krutrim & Sarvam AI: Indian AI Lab Interviews (ML / AI Engineer). Rounds, real reported questions and model answers, compiled from public candidate reports (sources at the end). Practice mode hides every answer: say your answer first, press Show answer, then rate yourself. Use 🔁 Review again to build your revision list and 🎲 Random question for a quick drill.

★ Must-know 145📝 Assignments
About this interview

Compiled from a public Krutrim AI Intern report (Bengaluru, university hiring, offer accepted: technical round with two Research Scientist-3 interviewers + HR) and public Sarvam AI ML Engineer role descriptions (speech and dubbing, vision, inference optimisation) with community discussion noting research-heavy, low-DSA interviews. The questions are paraphrased, and the model answers are written by EasyLearnSpace. Sarvam's specific round structure was not public in the sources we found, so treat that part as likely focus areas.

Krutrim (Ola) Sarvam AI
Role AI Intern; ML / Research roles ML Engineer (speech/dubbing, vision/VLM, Studio), 3–6 years' experience for mid roles
Rounds Technical (deep dive on your previous work) → HR (why Krutrim; a hard situation you worked through) Not public; candidates report a focus on depth and real work over LeetCode
Focus areas Your research or projects, LLM fundamentals Indic LLMs, speech (ASR/TTS), translation, evaluation of non-deterministic output, GPU inference optimisation
Outcome Offer accepted (rated easy) n/a

01Rounds at a glance

02Questions and model answers

Q1
Why do Indic languages need special handling in LLMs?
  • Tokenisation: English-centric BPE splits Devanagari and Tamil text into many tokens, which raises cost and latency and lowers quality. Indic-aware tokenisers (larger vocabularies, script-specific merges) cut the token count 2–4×.
  • Data scarcity and quality: less web text, so use translation-augmented, synthetic and curated data, and careful deduplication.
  • Code-mixing and transliteration: Hinglish in Latin script ("kal milte hain"). Train on romanised text too.
  • Evaluation: benchmarks such as IndicGLUE, IndicXTREME, FLORES for translation, and human evaluation by native speakers.
Q2
How would you evaluate a speech-to-speech dubbing pipeline (ASR → translation → TTS)?
  • Per component:
  • ASR: WER/CER.
  • Translation: BLEU/chrF/COMET plus human adequacy and fluency.
  • TTS: MOS, intelligibility (ASR on the TTS output), speaker-similarity cosine of voice embeddings.
  • End to end:
  • Lip-sync and timing alignment.
  • Preservation of emotion and prosody.
  • Naturalness ratings.
  • Error propagation analysis.
  • Non-determinism: fix seeds, sample several outputs, report distributions; use A/B preference tests against global baselines.
Q3
How do you optimise LLM inference on GPUs?
  • Serving engine: vLLM / TensorRT-LLM / SGLang with continuous batching, a paged KV cache, prefix caching.
  • Model: FP8/INT8/INT4 quantization (AWQ, GPTQ), GQA, speculative decoding (a small draft model), distillation.
  • Kernels: FlashAttention, fused ops, CUDA graphs.
  • Measure: time to first token, tokens per second per GPU, p99 latency, cost per 1M tokens; profile with Nsight. Tune batch size against latency targets.
Q4
Tell us about your previous work in depth (Krutrim technical round).

Answer (structure for research or engineering work):

  1. Problem and why it matters (one sentence).
  2. What existed before and its limitation.
  3. Your approach: the key idea, the architecture, the data, why it should work.
  4. Experiments: baselines, ablations, metrics, negative results.
  5. What you'd do next. - Research scientists will dig into ablations and failure cases, so be honest about what didn't work.
Q5
Why do you want to join an Indian AI lab rather than a big tech company?
  • "I want to build models for a billion users in their own languages, problems global labs under-serve: Indic speech, low-resource translation, voice-first interfaces."
  • "Smaller teams mean ownership across data, training and deployment."
  • Mention specific products (Sarvam's Indic models and voice stack; Krutrim's multilingual models and cloud) to show genuine interest.
Q6
How would you build an Indic-language RAG assistant for government-scheme questions?
  • Corpus: scheme PDFs in several languages → OCR as needed → chunk by section → multilingual embeddings (e.g. BGE-M3, LaBSE-style).
  • Query: speech input (ASR) in the user's language → retrieve across languages → answer in the user's language with an Indic-capable LLM → optional TTS.
  • Guardrails: cite the scheme clause; "I don't know" with a helpline fallback; simple vocabulary.
  • Evaluation: questions written by native speakers per language; faithfulness plus comprehension testing with real users.

03How to prepare

Sources

Created by Ajeetkumar