About the role
Responsibilities:
- Fine-tune foundation models for education-specific tasks: question answering, content generation, adaptive feedback, and curriculum alignment.
- Own end-to-end fine-tuning workflows: dataset curation, training runs, hyperparameter tuning, evaluation, and model versioning.
- Implement efficient fine-tuning methods (LoRA, QLoRA, DoRA, adapters) appropriate to available compute budgets.
- Build RLHF and preference optimisation pipelines: DPO, PPO, reward modelling for aligning models to learning outcomes.
- Optimise GPU training efficiency: DeepSpeed, FSDP, gradient checkpointing, mixed precision, multi-GPU setups.
- Evaluate fine-tuned models rigorously: perplexity, task-specific benchmarks, human eval, and regression testing.
- Build data pipelines for instruction tuning datasets including multilingual and Indic language data.
- Collaborate with AI platform teams on inference optimisation: quantisation, GGUF, ONNX export, vLLM serving.
- Assess new open-source model releases for domain applicability and adoption readiness.
- Define and track model performance metrics, evaluation benchmarks, and optimisation targets across all fine-tuned model versions.
- Work with open-source and sovereign LLMs, owning the full model adaptation lifecycle using proven, industry-standard frameworks.
Requirements:
- Strong experience in fine-tuning and optimising Large.
- Hands-on experience with LoRA, QLoRA, SFT, DPO, Language Models (LLMs), RLHF, or similar fine-tuning techniques.
- Proficiency in PyTorch and Hugging Face ecosystem (Transformers, PEFT, TRL).
- Experience with distributed and multi-GPU model training.
- Strong understanding of model performance, evaluation, and optimisation.
- Knowledge of DeepSpeed, Megatron-LM, and large- scale training frameworks.
- Understanding of AI/ML pipelines, data preparation, and model deployment.
- Experience working with open-source and sovereign LLMs; focus on standard, production-proven frameworks.
Preferred Skills:
- RAG integration: how fine-tuned models interact with retrieval systems and when to fine-tune vs retrieve.
- Inference optimisation: GGUF quantisation, ONNX export, TensorRT, vLLM serving.
- Vector databases: embedding generation, indexing, and hybrid search for downstream model use.
- Multilingual fine-tuning: Indic languages, code-switching, transliteration-aware tokenisation.
- Agentic AI: tool use, function calling, and fine-tuned model behaviour in agentic contexts.