高级软件工程师（机器学习）
Senior Software Engineer (Machine Learning)
一切以客户为中心！
荣登德勤科技Fast 500榜单和G2排行榜，Maropost提供统一的电商体验，满足客户所需，通过一个平台变革电商、零售、营销自动化、商品管理、客服和AI运营，该平台专为快速增长的企业设计。我们始终专注于客户的成功，以客户至上、紧迫感、卓越和创造力为动力，为全球5000+品牌提供支持，并朝着100000+的目标前进。
秉承上述以客户为中心的理念，我们赋能企业实现目标并与我们共同成长。如果你是一个主动推动者而非被动参与者，准备好产生重大影响并成为我们转型旅程的一部分，那么Maropost就是你的选择。
职位机会：
你将负责驱动商家产品发现的机器学习系统：个性化搜索排序、推荐组件、邮件推荐以及LLM购物助手。这是一个端到端的角色。你将训练模型，构建服务来部署它们，并确保两者在生产环境中运行。你不会把模型交给别人去处理。
团队规模小，系统真实：推荐结果在实时请求路径中重新排序，预算在200毫秒内，以商家转化率作为衡量标准。
你将负责以下工作：
- 设计、训练并发布推荐和排序模型——基于会话的嵌入、协同过滤、内容和视觉嵌入。
- 构建并维护低延迟的Python服务来提供这些模型。
- 运行每日重建模型制品的批处理管道，覆盖数百家商家。
- 设计并分析AB测试；根据证据决定哪些功能上线。
- 自主负责部署：容器、Kubernetes清单、自动扩展、仪表盘、警报。
- 扩展我们的LLM工作——购物助手和LLM辅助的商品目录增强——并有适当的评估支持。
你将为Maropost带来：
- 5年以上构建机器学习系统经验，至少3年模型在生产环境运行的经验。
- 精通Python（FastAPI或同等框架）。
- 向量数据库（Qdrant）。
- 其他ANN库——FAISS、Annoy、ScaNN。我们使用HNSWlib，但权衡方式是通用的。
- 有推荐系统或搜索排序的实际经验：隐式反馈、嵌入、近似最近邻搜索。
- 熟练使用SQL进行分析存储查询；有ClickHouse经验。
查看英文原文
Everything we do is for our customers!
Featured on Deloitte's Technology Fast 500 list and G2's leaderboard, Maropost offers a unified commerce experience that our customers need, transforming ecommerce, retail, marketing automation, merchandising, helpdesk and AI operations with one platform designed to scale for fast-growing businesses. With a relentless focus on our customers’ success, we are motivated by customer obsession, extreme urgency, excellence and resourcefulness to power 5,000+ global brands while we head to 100,000+.
Driven by the same customer-centric mentality as above, we empower businesses to achieve their goals and grow alongside us. If you're a driver and not passenger and are ready to make a significant impact and be part of our transformative journey, Maropost is the place for you.
The Opportunity:
You will own the machine learning that powers product discovery for our merchants: personalized search ranking, recommendation widgets, email recommendations, and an LLM shopping assistant. This is an end-to-end role. You will train the models, build the services that serve them, and keep both running in production. You will not hand a model over a wall to someone else.
The team is small and the systems are real: Recommendations are re-ranked in the live request path for storefronts, under a 200 ms budget, measured by merchant conversion.
What Youʼll Be Responsible For:
- Design, train and ship recommendation and ranking models - session-based embeddings, collaborative filtering, content and visual embeddings.
- Build and operate the Python services that serve them at low latency.
- Run the batch pipelines that rebuild model artifacts daily across hundreds of merchants.
- Design and read AB tests; decide what ships on the evidence.
- Own your deployments: containers, Kubernetes manifests, autoscaling, dashboards, alerts.
- Extend our LLM work - a shopping assistant and LLM-assisted catalogue enrichment - with proper evaluation behind it.
What Youʼll Bring to Maropost:
- 5+ years building machine learning systems, with at least 3 years of models serving live production traffic.
- Strong Python,(FastAPI or equivalent).
- Vector databases (Qdrant).
- Other ANN libraries - FAISS, Annoy, ScaNN. HNSWlib is what we run, but the trade-offs transfer.
- Hands-on recommender systems or search ranking experience: implicit feedback, embeddings, approximate nearest neighbour search.
- Solid SQL against analytical stores; ClickHouse experience is a plus.
- Comfortable with Docker and Kubernetes, and willing to own the deployment of your own work.
- Experience running AB tests and reporting results you did not like.
- Streaming systems (Pulsar, Kafka, Flink)
- You exemplify Maropost’s Values:
Customer Obsessed
Extreme Urgency
Excellence
Resourceful
Bonus Points:
- PyTorch, sentence-transformers and CLIP, or computer vision applied to product imagery.
- Gradient boosting for ranking (XGBoost, LightGBM or CatBoost). We run a legacy XGBoost autocomplete ranker and expect to revisit learned ranking.
- Scikit-learn and scipy for lightweight classifiers and experiment statistics.
- LLM application work with evaluation harnesses, tool calling, and cost and latency tuning.
- E-commerce, search relevance, or marketplace background.
- Python async experience
- Experience around search
Our stack:
Python 3.11–3.13, FastAPI, gensim, PyTorch, sentence-transformers and CLIP, HuggingFace transformers, XGBoost, scikit-learn, scipy, pandas, numpy, HNSWlib, Qdrant, ClickHouse, Redis, MySQL, Postgres, Pulsar, Docker, Kubernetes on GCP, ArgoCD, CircleCI, Grafana, Sentry, and Gemini on Vertex via pydantic-ai.
Message from the Founders: Maropost is looking for builders - people who want to drive our business forward at all costs in order to achieve the goals we have both short and long term for the results and outcomes that that will bring to us all.
If that isn't for you that’s ok, for those of you that it is please get in touch with us!
Originally posted on Himalayas