Data Scientist (AI & ML)

Data Scientist (AI & ML)

Role Summary

The Data Scientist will design, develop and evaluate statistical, machine learning and AI solutions, including LLM applications, retrieval systems and deep learning models for text, image, speech and document data. The role requires sound judgement in selecting the appropriate technique, rigorous validation of predictive and generative models, and delivery of production-quality code.

Educational Qualification

• B.Tech./B.E., M.Tech./M.S. or M.Sc. in Computer Science, Statistics, Mathematics, Data Science, AI/ML or a related quantitative discipline (Must have)
• Advanced degree (M.S./M.Tech./Ph.D.) or published applied work in machine learning, NLP, computer vision or statistics is highly desirable
• Certifications are not mandatory; candidates with strong verifiable work (Kaggle Expert level or above, published research or significant open-source contributions) will be considered in lieu of formal qualification

Experience

• 4+ years in data science, machine learning or applied AI (6+ years preferred); minimum 3 years developing models or AI systems deployed in production or used in operational decisions
• Demonstrable end-to-end ownership of at least two models or AI systems, with measured outcomes
• Hands-on experience with at least one LLM or deep learning application deployed for end users, with documented evaluation
• Experience in at least two of the following: fraud or anomaly detection, forecasting, risk scoring, NLP and LLM applications, computer vision or document AI, speech, and causal inference
• Experience with large, heterogeneous datasets, including linkage of records across sources
• Experience in a regulated sector (financial services, healthcare, telecommunications or the public sector) or with Indian-language data is a plus).

Key Responsibilities

Problem Formulation

• Translate business requirements into well-defined analytical problems with a clear decision, baseline and evaluation metric
• Select the appropriate approach (rules-based, classical machine learning, deep learning, a fine-tuned model or an LLM) based on accuracy, cost, latency and explainability, and document the rationale
• Assess data feasibility, including label quality, leakage and bias, and advise where the data is insufficient for the objective

Machine Learning & Statistical Modelling

• Develop and validate models for classification, regression, ranking and forecasting, starting from robust baselines
• Develop fraud and anomaly detection models for imbalanced settings with limited labelled data
• Generate labels through weak supervision, active learning and LLM-assisted labelling, with human validation

Generative AI & Deep Learning

• Develop LLM-based solutions, including retrieval-augmented generation, document extraction, summarisation and tool-using agents
• Design retrieval components (chunking, embeddings, hybrid search and re-ranking), and adapt models through prompting, parameter-efficient fine-tuning and distillation
• Develop deep learning solutions for NLP, document understanding (OCR and vision-language models), vision and speech, including Indian-language and code-mixed data
• Optimise cost and latency, including the choice between open-weight and API-based models

Evaluation & Statistical Rigour

• Validate models using leakage-free schemes, and quantify uncertainty through prediction intervals, calibration and conformal prediction
• Evaluate generative AI systems through golden datasets, separate retrieval and generation metrics, hallucination checks, and LLM-based evaluation calibrated against human labels
• Conduct error and subgroup analysis across cohorts, languages and input types
• Design experiments and apply quasi-experimental methods to assess whether deployed models have improved outcomes

Production & Monitoring

• Develop tested, peer-reviewed, production-quality Python code and ensure all experiments are reproducible
• Hand over models to the MLOps team with defined monitoring metrics, drift thresholds and retraining triggers
• Monitor deployed systems for quality, drift, latency and cost, and incorporate failures into evaluation sets

Responsible AI & Privacy

• Measure and mitigate performance disparities across cohorts, and provide explanations for decisions affecting individuals
• Test LLM systems for prompt injection, jailbreaks and leakage of personal data, and design guardrails in coordination with security specialists
• Apply anonymisation and data minimisation in compliance with applicable data protection law, and prepare model cards and evaluation recordsCommunication & Mentoring
• Prepare clear reports for non-technical stakeholders, including model limitations and uncertainty
• Review the work of junior data scientists, and support the review of vendor model deliverables

Technical Competencies

• Statistics & Causal Inference: generalised linear and hierarchical models, Bayesian inference, hypothesis testing, bootstrap methods; A/B testing, difference-in-differences, synthetic control, regression discontinuity (DoWhy, EconML)
• Machine Learning: XGBoost, LightGBM and CatBoost; tabular foundation models (TabPFN); anomaly detection, clustering and graph methods; forecasting with statistical models and time-series foundation models (Chronos-2, TimesFM)
• Generative AI & LLMs: prompt design, structured outputs and tool calling; retrieval-augmented generation with hybrid search and re-ranking; vector databases (pgvector, Qdrant, Milvus or Weaviate) and FAISS; open-weight and API-based models; agent frameworks (LangGraph, LlamaIndex or equivalent) and the Model Context Protocol (MCP)
• Model Adaptation & Serving: LoRA or QLoRA fine-tuning and preference optimisation (DPO, GRPO) using Hugging Face TRL, PEFT or Unsloth; distillation; quantisation; efficient serving (vLLM or SGLang)
• Deep Learning & Document AI: PyTorch and Hugging Face Transformers; vision-language models for document parsing (such as PaddleOCR-VL, Docling or the Qwen-VL family); vision models for classification and detection; speech recognition (IndicConformer, Whisper or newer open models)
• Indic Language AI: AI4Bharat models (IndicConformer speech recognition for 22 Indian languages, IndicTrans2 translation, IndicBERT, Indic Parler-TTS), Sarvam models and Bhashini services; handling of code-mixed and transliterated text
• Evaluation & Calibration: RAGAS, DeepEval, promptfoo or Inspect; retrieval metrics (recall at k, MRR, nDCG); calibration and conformal prediction; SHAP and Fairlearn
• Data, Engineering & Observability: advanced SQL, PySpark, Polars or DuckDB; Python with pytest, Git, MLflow, DVC and Docker; LLM tracing and observability (Langfuse or Arize Phoenix, OpenTelemetry)
• AI Safety & Privacy: prompt-injection and data-leakage testing, guardrail frameworks (NeMo Guardrails, Llama Guard), anonymisation, working knowledge of differential privacy
• Communication: Matplotlib, Plotly, Power BI or Tableau; explaining model limitations to non-technical audiences