Machine Learning Expert for Recommendation&Ranking Systems

About the role Freeleaps runs a freelance marketplace, and magicleaps a job-search product. Both are served by our own recommendation service, written in Python: embedding retrieval with pgvector, a LightGBM LambdaRank reranker, daily training, and a promotion gate that decides whether a new model replaces the one in production. The system works end to end. What limits it today is thin, noisy data and an evaluation process that cannot yet tell a real gain from noise. We are looking for an expert who thinks like an ML practitioner: you ask what the data can and cannot prove before changing anything, and you measure the change afterwards. There are many problems to solve, from ranking quality to the architecture of the ML system, and you will take them on one assignment at a time. Job responsibilities As our machine learning expert, you will: • Assess the matching pipeline and rank what limits accuracy, with evidence. • Fix training-data problems: label construction from clicks and applications, de-duplication, grouping and reproducible runs. • Build offline evaluation: rolling time windows, bootstrap confidence intervals, and a promotion gate that decides on intervals. • Improve ranking features and measure retrieval recall. • Shape the ML architecture: training and serving pipelines, identical features in training and serving, model registry and rollout, and evaluation infrastructure. • Take on future problems such as Chinese and English text matching, cold start and two-sided matching. Qualifications Required • Expert-level applied machine learning for ranking or recommendation: learning to rank, implicit feedback, position bias and leakage. • Experience designing ML systems end to end: data pipelines, training, serving, monitoring and rollout. • Strong statistics for small, noisy data: bootstrap intervals, effect sizes, and knowing when a result is inconclusive. • Solid Python engineering: unit tests, reproducible experiments, and pull requests a reviewer can follow. • Clear written English. Preferred • Experience with LightGBM or XGBoost ranking. • Experience with vector search and sentence embeddings. • Experience with Chinese and English text. • Experience with two-sided marketplaces. Engagement and compensation • Part-time and remote, on hours you choose. • Each assignment has a fixed price and is paid when its milestone passes review. • Your first assignment is an assessment of our recommendation pipeline plus one verified change: about 10 to 25 hours, for USD 300 to 1,000. • You receive read access to the code and a pseudonymised data snapshot. You never need access to production. • We review every pull request within two business days.