Want to know if this job is worth applying to?
深圳市, 广东省, 中国
Onsite
Own recommendation and personalization end-to-end: Build and iterate the ML systems behind ShopBack's personalized shopping experience, recommendations, ranking, user modeling, and CRM intelligence. Own the change end-to-end — dataset, training, offline eval, A/B, rollout — and be measured on offline and online metrics.
Blend classic ML with modern LLM techniques: Apply fine-tuned open-source models, embedding models, and LLM-based models where they beat classic methods, and know when they don't.
Ship with evidence: Build evaluation sets and experiment harnesses before shipping models; foster a fast-paced, high-iteration experimentation culture (A/B, interleaving, causal reads).
Raise the team: Mentor our ML engineers on modern recommendation and LLM practice; your success includes the team's growth, not just your own output.
Metrics driven: Understand the business and product metrics behind personalization, and drive efforts that move them.
Handle ambiguity: Navigate loosely defined problems effectively, with or without dedicated Product Manager support.
Collaboration: Work closely with product, ops, and CRM stakeholders to set and achieve optimal outcomes.
Has shipped and iterated recommendation / personalization / ranking or search systems serving millions of users, and can walk through what moved the metrics, what didn't, and why (typically 2+ years of industrial ML experience).
Strong grounding in retrieval and ranking modeling, embeddings, and online experimentation.
Hands-on fine-tuning of open-source models/LLMs (SFT / LoRA / DPO) applied to ranking, personalization, or user modeling, and the judgment of when classic methods win.
Builds evaluation sets and harnesses as a default step, not an afterthought.
Understands dataset licensing and provenance for commercial use.
Solid MLOps fundamentals: data pipelines, productionisation, monitoring, and GPU cost awareness.
Comfortable in a batch data stack — Spark or equivalent, a scheduler(e.g. Airflow), cloud training and serving (e.g. AWS SageMaker). You’d own the model through ingestion, training, serving, monitoring, retraining and rollback.
Strong Python and PyTorch; familiarity with the Hugging Face ecosystem (transformers / PEFT / TRL) and modern inference stacks (e.g. vLLM) is a plus.
Uses agentic AI tools as a daily driver for engineering work, and can show how they changed your workflow.
Education in a quantitative field such as Computer Science, Statistics, or Mathematics, or equivalent practical depth.
Strong desire to solve tough problems with scientific rigour at scale, and to get results early and iterate.