hands-on Data Scientist (NLP and Generative AI to)

Job Summary

We are looking for a hands-on Data Scientist with deep expertise in NLP and Generative AI to help

build and refine the intelligence behind our agentic AI systems. You will be responsible for fine-

tuning, prompt engineering, and evaluating LLMs that power our digital workers across real-world

workflows.

Years of Experience 3 - 6 Years

Budget 18 LPA to 24 LPA

• Location Chennai

Immediate to 30 days

Key Responsibilities

·     Fine-tune and evaluate LLMs (e.g., Mistral, LLaMA, Qwen) using frameworks like Unsloth,

HuggingFace, and DeepSpeed

·     Develop high-quality prompts and RAG pipelines for few-shot and zero-shot performance

·     Analyze and curate domain-specific text datasets for training and evaluation

·     Conduct performance and safety evaluation of fine-tuned models

·     Collaborate with engineering teams to integrate models into agentic workflows

·     Stay up to date with the latest in open-source LLMs and GenAI tools, and rapidly prototype

experiments

·     Apply efficient training and inference techniques (LoRA, QLoRA, quantization, etc.)

Requirements

·     3+ years of experience in Natural Language Processing (NLP) and machine learning applied to

text

·     Strong coding skills in python

·     Hands-on experience fine-tuning LLMs (e.g., LLaMA, Mistral, Falcon, Qwen) using frameworks

like Unsloth, HuggingFace Transformers, PEFT, LoRA, QLoRA, bitsandbytes

·     Proficient in PyTorch (preferred) or TensorFlow, with experience in writing custom

training/evaluation loops

·     Experience in dataset preparation, tokenization (e.g., Tokenizer, tokenizers), and formatting

for instruction tuning (ChatML, Alpaca, ShareGPT formats)

·     Familiarity with retrieval-augmented generation (RAG) using FAISS, Chroma, Weaviate,

or Qdrant

·     Strong knowledge of prompt engineering, few-shot/zero-shot learning, chain-of-thought

prompting, and function-calling patterns

·     Exposure to agentic AI frameworks like CrewAI, Phidata, LangChain, LlamaIndex, or AutoGen

·     Experience with GPU-accelerated training/inference and libraries

like DeepSpeed, Accelerate, Flash Attention, Transformers v2, etc.

·     Solid understanding of LLM evaluation metrics (BLEU, ROUGE, perplexity, pass@k) and safety-

related metrics (toxicity, bias)

·     Ability to work with open-source checkpoints and formats (e.g., safetensors, GGUF, HF

Hub, GPTQ, ExLlama)

·     Comfortable with containerized environments (Docker) and scripting for training

pipelines, data curation, or evaluation workflows

Nice to Haves

·     Experience in Linux (Ubuntu)

·     Terminal/Bash Scripting