Please read JD carefully before applying. You will only be shortlisted if your resume if aligned with the JD.

We are looking for a Data Scientist who writes production code every day, ships models that receive real traffic, and is as comfortable debugging a RAG pipeline as tuning an XGBoost model for pricing signals.

This is a hands-on individual contributor role. You report to the CTO and work in collaboration with the ML Engineer.

The ideal candidate brings a strong classical ML foundation and has meaningfully expanded their stack into GenAI engineering : RAG, embeddings, LLM tool use, and prompt evaluation. You are not waiting to be assigned problems; you identify gaps in the platform, propose solutions, and ship them independently.

We are not looking for someone focused on people management, strategy presentations, or org-building. We want someone who wants to go deep, stay technical, and own the full ML lifecycle end-to-end across a genuinely complex, high-stakes domain.

What You'll Do :

- Build and own production ML models for pricing signal detection, demand forecasting, lead time prediction, lifecycle risk scoring, and inventory age risk primarily using XGBoost, LightGBM, CatBoost, and time-series methods (ARIMA, Prophet, temporal Transformers where warranted).

- Own the full ML lifecycle end-to-end : feature engineering, model training, offline evaluation, A/B testing, production deployment, drift monitoring, and scheduled retraining pipelines.

- Make principled decisions on when a gradient-boosted model or a simple statistical heuristic outperforms a large LLM; optimising for cost, latency, and explainability at every layer.

- Build SHAP-based interpretability outputs and business-readable model explanations so that pricing and risk models are trusted, auditable, and adoptable by product and operations teams.

- Design and execute A/B tests and experimentation frameworks to measure the business impact of each model within its first sprint in production.

- Build RAG pipelines for datasheet extraction, BOM parsing, RFQ drafting, and supplier communication; grounding LLMs in proprietary component intelligence data.

- Design and implement embedding strategies: select embedding models, define chunking approaches, implement hybrid dense-sparse retrieval (Qdrant, Weaviate, pgvector), and build reranking layers tuned per use case.

- Integrate frontier model APIs (OpenAI, Anthropic, Gemini) and self-hosted open-source models (Llama 3, Mistral, Qwen via HuggingFace) based on cost, latency, and accuracy benchmarks; with documented trade-off rationale.

- Build and maintain prompt engineering systems that are version-controlled, systematically evaluated, and reproducible, not ad-hoc prompts left in notebooks.

- Implement LLM evaluation pipelines using RAGAS or equivalent frameworks to track retrieval quality, answer relevance, and hallucination rates across all RAG systems.

- Design, build, and maintain feature pipelines using dbt, Pandas, and Polars, transforming raw procurement and component data into high-quality training and serving features.

- Work closely with the data stack (Postgres, Redshift, S3) to build efficient, reproducible feature stores and training datasets with clear lineage.

- Implement data quality checks and validation layers for ML inputs, catching distribution shifts at the data layer before they surface in model behaviour.

- Build and document feature registries so that features engineered for one model are discoverable and reusable across the platform.

- Apply leakage-free cross-validation design : proper temporal splits, group k-fold, and nested CV strategies appropriate to each prediction task.

- Own experiment tracking using MLflow (self-hosted), every experiment logged, every result reproducible, every model version documented with a clear performance baseline.

- Containerise model inference services using Docker and deploy on AWS (SageMaker, Lambda, ECS) with clear latency SLAs, runbooks, and rollback procedures.

- Implement production monitoring using Evidently AI or equivalent: data drift, model performance drift, and business KPI alignment, automated alerts before humans notice degradation.

- Maintain clean CI/CD pipelines (GitHub Actions) for model deployment: code review, automated unit and integration tests, staging promotion, and production release gates.

- Actively use AI-assisted development tools (Cursor, GitHub Copilot, Windsurf) to compound your own output velocity, vibe coding is a first-class practice here, not a side habit.

- Build Python-based tools, function schemas, and data connectors that the Director uses in multi-agent orchestration workflows built on LangGraph, CrewAI, or AutoGen.

- Design and expose MCP-compatible tool endpoints for model-serving, data retrieval, and scoring functions, making ML model outputs composable and consumable by AI agents across the platform.

- Support agentic workflow testing : write evaluation suites, simulate tool misuse scenarios, and build regression tests for agent-facing ML components.

- Stay current with the agentic AI and MCP ecosystem, read agent traces, review LangGraph state graphs, and flag failure modes in agent-model integration before they reach production.

Must-Have :

- You ship code daily. Your GitHub is active. Python is your primary language. You debug model traces, read stack traces, and review your own PRs, you do not wait for an engineering team to implement your ideas.

- 6 - 12 years of hands-on experience in data science, ML engineering, or software engineering with a clear arc toward production AI/ML systems.

- At least 3 years owning production ML systems, not research prototypes. Models that received real traffic, failed in real ways, and were debugged, retrained, and monitored over time.

- Strong classical ML proficiency : XGBoost, LightGBM, CatBoost for tabular tasks; time-series forecasting; clustering and unsupervised methods; SHAP-based interpretability. You know when gradient boosting beats a neural network and have the benchmark to prove it.

- Feature engineering depth: encoding strategies, temporal and lag features, interaction terms, domain-specific transforms, and leakage-free cross-validation design. This is still more impactful than model choice on most business problems.

- Practical GenAI engineering : you have built RAG pipelines in production. You understand chunking strategy trade-offs, embedding model selection, retrieval quality metrics, and reranking layer design.

- Vector database fluency : hands-on experience with Qdrant, Weaviate, ChromaDB, or pgvector- index types (HNSW, IVF), metadata filtering, and hybrid dense-sparse search.

- LLM orchestration experience: LangChain or LlamaIndex at production scale, chains, retrieval pipelines, and version-controlled prompt management.

- Prompt engineering as a discipline : version-controlled, systematically evaluated prompts. Chain-of-thought, few-shot, and XML-tagged system prompts. Not ad-hoc iteration in a notebook.

- MLOps hygiene : MLflow or W&B for experiment tracking, Docker for containerisation, AWS basics (SageMaker, S3, Lambda), and model monitoring in production.

- HuggingFace ecosystem: model hub, transformers library, datasets, sentence-transformers. You can benchmark and select an embedding model for a specific retrieval task, not just call from_pretrained().

Preferred :

- Hands-on experience with MCP (Model Context Protocol), building or consuming MCP-compatible tool endpoints for AI agent integration.

- Fine-tuning experience: LoRA, QLoRA, PEFT on open-source LLMs (Llama, Mistral, Qwen) for domain adaptation or task-specific performance improvement.

- LLM evaluation frameworks: RAGAS, LLM-as-judge patterns, or custom eval harness design for RAG pipeline quality measurement.

- Experience contributing to agentic workflows, building tool schemas, function signatures, or supporting multi-agent state machine design in LangGraph or equivalent.

- Exposure to supply chain, procurement, logistics, or industrial B2B data, pricing models, demand forecasting, or lifecycle risk in physical goods contexts.

- Familiarity with Ollama or vLLM for local inference and cost benchmarking against cloud API alternatives.