Company Interview Questions · IT Services

LTIMindtree & HCLTech: GenAI / ML Engineer

❓ 10 questions🏢 LTIMindtree & HCLTech🆕 Recent reports

LTIMindtree & HCLTech: GenAI / ML Engineer Interview Questions (2025). 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 process guide for the LTIMindtree GenAI Engineer role and a public HCLTech ML / GenAI candidate report (Bengaluru, July 2025). The questions are paraphrased, and the model answers are written by EasyLearnSpace.

	LTIMindtree	HCLTech
Role	GenAI Engineer	ML / Gen AI Engineer
Rounds	Up to 6: screening → fundamentals → technical deep dive → architecture → managerial → HR	A single expert panel
Timeline	About 4–6 weeks	Not stated
Style	Fundamentals + enterprise system design	Scenario-heavy: an NLP coding question plus several scenarios
Outcome reported	Not stated	Not selected; the panel was described as very knowledgeable and the interview as difficult

01Rounds at a glance

LTIMindtree: Transformers vs RNNs, data leakage, SFT vs RLHF, evaluating generative models, an enterprise RAG pipeline, latency, versioning and drift, API vs open source, protecting sensitive data, plus LoRA/PEFT and CI/CD for LLM apps.
HCLTech: which NLP libraries you would use and how; then scenarios to analyse live, so think aloud.

02Questions and model answers

What is data leakage and how do you prevent it?

Definition: information from outside the training data, often the target or the future, gets into training. Validation scores then look great while production performs badly.
Common causes:
Scaling or imputing on the full dataset before the split.
Target encoding without folds.
Random splits on time-series data.
Duplicate rows that fall into both train and test.
Features that are only known after the outcome.
Prevention:
Split first.
Fit preprocessing inside a Pipeline per CV fold.
Use time-based or group-based splits.
Deduplicate.
Review every feature for "would I know this at prediction time?"

What is the difference between supervised fine-tuning (SFT) and RLHF? Where does DPO fit?

SFT: train on (prompt → ideal response) pairs with next-token loss. It teaches format, style and how to follow instructions.
RLHF: 1. Collect human preferences (response A is better than B). 2. Train a reward model on them. 3. Optimise the policy with PPO to maximise reward, with a KL penalty to stay close to the SFT model.
It aligns the model to helpfulness and harmlessness.
DPO: optimises directly on preference pairs with a simple classification-style loss, with no reward model and no RL loop. It is simpler and more stable, and widely used now.
Typical pipeline: pre-training → SFT → preference tuning (RLHF, DPO or ORPO).

How do you evaluate generative models beyond perplexity?

Perplexity only measures how well the model predicts text. It doesn't tell you whether the answer is correct, helpful or safe.
Reference-based:
Exact match or F1 for QA
BLEU or ROUGE for translation and summaries (limited)
BERTScore (semantic)
Reference-free: LLM-as-judge with a rubric, pairwise preference (win rate), faithfulness to the context, and RAGAS for RAG.
Task benchmarks: MMLU, GSM8K, HumanEval, plus a domain-specific test set.
Human evaluation on a sample, plus safety and toxicity tests, latency and cost.

How do you reduce LLM inference latency?

Model side:
Use a smaller or distilled model.
Quantize (INT8 / 4-bit).
Speculative decoding.
Serving:
vLLM or TGI with continuous batching and PagedAttention for the KV cache.
Tensor parallelism.
Optimised kernels (FlashAttention).
Application:
Stream tokens.
Cache responses.
Shorten prompts and outputs.
Use prompt or prefix caching.
Run retrieval and tools in parallel.
Route simple queries to small models.
Measure time to first token and tokens per second separately; they have different fixes.

How do you handle model versioning and drift for an LLM application?

Version everything:
Model name and version
Prompt templates (in git)
Embedding model
Index snapshot
Evaluation set
Record them all in each trace.
Kinds of drift:
Data/query drift: users ask new things. Monitor the embedding distribution and topic clusters.
Knowledge drift: documents go out of date. Use a refresh pipeline and freshness metadata.
Provider drift: the hosted model is updated. Pin versions and re-run the evaluation before upgrading.
Respond: run the evaluation set nightly or weekly, alert on drops in quality or feedback, and roll back via a model or prompt registry.

How do you protect sensitive data when fine-tuning?

Minimise: use only the fields you need, and detect and mask PII (Presidio, regex + NER) before training.
Environment: train in a private VPC or VNet, keep encryption at rest and in transit, and enforce strict IAM. Use a cloud provider's no-retention fine-tuning or self-host.
Memorisation risk: deduplicate, train for fewer epochs, consider differential privacy for highly sensitive data, and test for regurgitation of training records.
Governance: data-use approval, lineage, model cards, and access control on the resulting model.

What is LoRA/PEFT and why use it?

LoRA: freeze the base weights W and learn a low-rank update ΔW = B·A (rank r = 4 to 64) only on selected layers, usually the attention projections.
Benefits:
Trains under 1% of the parameters.
Needs far less GPU memory.
Produces small adapter files, one per task, that can be swapped.
Avoids catastrophic forgetting better than a full fine-tune.
QLoRA: LoRA on a 4-bit quantized base, so a 7B to 13B model can be fine-tuned on one GPU.
Other PEFT methods: prefix or prompt tuning, IA³, adapters.

What does CI/CD look like for an LLM application?

CI (on each pull request):
Unit tests
Linting
Prompt and chain tests with mocked LLMs
An evaluation gate: run the golden set and fail if faithfulness or accuracy drops more than X%
Security scans
CD:
Build the container.
Deploy to staging.
Run smoke and evaluation tests.
Canary or blue-green release to production.
Automatic rollback on alerts.
Artefacts: prompts, configs and model versions in a registry; infrastructure as code (Terraform or Bicep).

HCLTech: which NLP libraries would you use for a text classification task, and how?

Quick baseline: scikit-learn TfidfVectorizer + LogisticRegression in a Pipeline, evaluated with stratified CV.
Preprocessing:
spaCy for tokenisation, lemmas and NER.
NLTK for stop words and stemming in classic pipelines.
re for cleaning.
Better accuracy: Hugging Face transformers, fine-tuning DistilBERT or RoBERTa with Trainer, or sentence-transformer embeddings with a classifier.
LLM option: zero- or few-shot classification with structured output, useful when labels are scarce.
Explain why: start simple, measure, then increase complexity only if needed.

Q10

HCLTech scenario: the chatbot gives correct answers in testing but users complain in production. What do you do?

Collect evidence: pull production traces and look for negative-feedback clusters. Are failures tied to a topic, a language or long chats?
Compare distributions: production questions are often messier than the test set (typos, Hinglish, multi-part questions).
Find the root cause: - Retrieval misses (low scores)? - Stale documents? - Lost context in long chats? - Latency timeouts?
Fix and prevent: - Add real queries to the evaluation set. - Query rewriting. - Better chunking. - Memory summarisation. - Monitor feedback by topic.
Communicate progress to stakeholders with metrics.

03How to prepare

LTIMindtree: many rounds, so keep stamina for architecture and managerial rounds, and know PEFT and CI/CD for LLMs.
HCLTech: practise scenario questions aloud with a clear structure (clarify → hypothesise → check → fix → prevent).
For both: prepare one enterprise RAG design with security and cost.

Sources

Created by Ajeetkumar