About the role
Responsibilities:
- Design, develop, and deploy AI/ML models and pipelines in production environments.
- Implement Retrieval-Augmented Generation (RAG) architectures and agentic AI workflows.
- Fine-tune and optimize LLMs for domain-specific use cases using RLHF, LoRA, and QLoRA.
- Build robust prompt engineering frameworks and evaluation pipelines.
- Integrate LLM APIs (OpenAI, Claude, Gemini) and open-source models into product features.
- Develop and maintain vector search infrastructure and embedding pipelines.
- Collaborate with architects, backend engineers, and product teams on AI feature delivery.
- Monitor model performance, conduct A/B testing, and iterate based on metrics.
- Implement guardrails, safety layers, and hallucination-mitigation strategies.
- Contribute to MLOps practices: model versioning, deployment pipelines, and monitoring.
Requirements:
- Strong expertise in Python, with deep knowledge of AI/ML libraries (PyTorch, TensorFlow, Hugging Face Transformers).
- Hands-on experience with LLM APIs and prompt engineering techniques (CoT, few-shot, ReAct).
- Experience with RAG systems, embedding models (text-embedding-3, BGE, Cohere), and vector stores.
- Knowledge of agentic frameworks: LangChain, LlamaIndex, AutoGen, CrewAI, or Semantic Kernel.
- Familiarity with fine-tuning techniques: LoRA, QLoRA, PEFT, instruction tuning.
- Experience deploying models on cloud platforms (AWS SageMaker, GCP Vertex AI, and Azure ML).
- Understanding of data preprocessing, feature engineering, and model evaluation metrics.
- Proficiency with MLOps tools: MLflow, DVC, Weights and Biases, BentoML
- Experience with containerization and orchestration: Docker, Kubernetes.
- Strong debugging and experimentation skills with Jupyter, FastAPI, Streamlit.
Nice-to-have:
- Experience with multi-modal models (vision-language models, Whisper, DALL-E); published papers or Kaggle/competition achievements.
- Exposure to speech AI, computer vision, or NLP specializations.
- Knowledge of responsible AI, fairness metrics, and bias.