Generative AI engineer in retail

We are seeking a Generative AI Engineer to own how the platform's language models are improved — from fine-tuning and evaluation to deployment. You will partner with Back-End engineers who handle the serving infrastructure, and collaborate with Data Analysts and business stakeholders who define what a «good» recommendation actually looks like.

• Run supervised fine-tuning and parameter-efficient adaptation (LoRA/QLoRA, adapters) and preference-based alignment (DPO, ORPO, RLHF-style methods) to shape tone, output structure, and domain reasoning for retail next-best-action and content-compliance tasks
• Own the dataset lifecycle — sourcing and curating examples from platform data and expert feedback, designing annotation and review workflows, deduplication, and decontamination against eval sets
• Generate and filter synthetic data, and version every dataset used in training
• Reduce cost and latency through distillation into smaller models, quantization, batching and caching strategy, and model routing — while holding quality above an agreed threshold
• Package and deploy tuned models on Vertex AI (or equivalent), and manage model and prompt versioning
• Monitor drift and quality decay in production, and define the retraining trigger and rollback path
• Implement and tune safety, scope, and PII controls appropriate to enterprise commercial data, including refusal behavior and output validation
• Partner with commercial teams to convert qualitative expectations ("this recommendation reads wrong") into training signal and measurable eval criteria

• 5+ years of hands-on experience with Large Language Models (LLM) fine-tuning — SFT and PEFT/LoRA at minimum, ideally preference optimization (DPO or similar)
• At least 1 year of relevant leadership experience
• Proficiency in Hugging Face Transformers/PEFT/TRL, Axolotl, or Unsloth, or managed tuning on Vertex AI / Bedrock / OpenAI
• Understanding of training mechanics, including tokenization, context length and packing, and learning-rate and hyperparameter choices
• Knowledge of catastrophic forgetting, overfitting on small domain sets, and GPU/TPU memory constraints
• Expertise in LLM evaluation — designing eval sets and metrics for generative tasks, with healthy skepticism about benchmark scores that don't reflect user outcomes
• Skills in data curation with strong instinct for dataset quality over quantity — labeling guidelines, inter-annotator agreement, and leakage control
• Familiarity with synthetic data pitfalls and filtering strategies
• Background in LLMOps, including model and prompt versioning, drift monitoring, and retraining lifecycle management
• Competency in Google Cloud Platform (Vertex AI, GCS, Cloud Run) and Git-based workflows
• Sufficient Python skills to work in notebooks, training scripts, and eval pipelines, and to adapt existing framework code
• Excellent command of written and spoken English (B2+ level)
• Будет плюсом: Experience with Reinforcement Learning from Human Feedback (RLHF) and preference-based alignment methods; Familiarity with distillation and quantization techniques (GPTQ, AWQ, GGUF, bitsandbytes) and their quality cost; Understanding of inference stacks such as vLLM, TGI, and Vertex AI endpoints, including streaming and latency/throughput implications; Knowledge of LLM security, including safety guardrails, scope controls, and PII handling for enterprise commercial data

• Условий в вакансии нет