Teleshop (HK) Ltd is a global sourcing and export company based in Hong Kong, supplying supermarkets, retail chains, and brands across Europe, Asia, and the Americas. We specialise in home appliances, beauty appliances, cleaning tools, and pet-care products, and our operations cover supplier discovery, product development, quality control, certifications, and logistics management.
We hold more than twenty thousand historical B2B export transactions with their real commercial outcomes attached, including the disputes, delays, and losses that followed. This is an unusual dataset in a domain almost nobody has modelled seriously. We are building a dedicated R&D function to work on it, applying machine learning to the specific problems of B2B export commerce, where conventional consumer-facing methods do not transfer.
Role Description
This is a full-time, on-site research and engineering role focused on estimating the risk-adjusted realised value of a B2B export deal. Not just whether a buyer is likely to want a product, but whether the resulting deal will actually be profitable once freight, duties, financing, disputes, returns, late payment, and other real-world losses are accounted for. These loss events are rare, highly variable in size, unfold over months, and are recorded only informally across correspondence and operational records, which makes estimating them a genuine technical problem rather than a reporting exercise.
You will design, build, and ship the models that turn this sparse, messy signal into reliable, decision-relevant value estimates, and validate whether ranking recommendations by risk-adjusted value improves real commercial outcomes over ranking by buyer relevance alone. This is a hands-on role: you will write the data-processing code, implement the models, build the evaluation harness, and be responsible for clean, reproducible, well-tested code end to end.
The work spans the full lifecycle: constructing supervised signal from informal records, building probabilistic loss models under sparsity, correcting for the statistical biases in historical sales data, and validating the end-to-end system through controlled evaluation. The first six months focus on label construction and the loss model. Factor decomposition, the re-ranking module, and end-to-end evaluation follow from there.
You will work alongside our ML Research Lead and with an academic advisor at the University of Hong Kong, with support for publishing work that merits it.
Key Responsibilities
Design procedures to infer loss events, their severity, and a confidence score for each from informal records (correspondence, accounting and shipping records), and run structured expert-validation of the resulting labels.
Develop probabilistic models of loss frequency and severity under sparse and incomplete data.
Train and evaluate models end to end in Python, from baselines through to final evaluation.
Build offline evaluation benchmarks and correct for selection bias in historically logged decisions (off-policy / counterfactual estimation, econometric selection models).
Write clean, well-structured, reproducible code: data pipelines, model implementations, and an evaluation harness that others can run and extend.
Design and run controlled evaluations, define their metrics and guardrails, and analyse results.
Produce clear technical documentation, ablation analyses, and reports.
Essential Qualifications
Bachelor's, Master's, or doctoral degree in a STEM-related discipline (Computer Science, Statistics, Data Science, Operations Research, Applied Mathematics, Econometrics, or related).
Strong programming ability in Python. You should write clean, idiomatic, well-organised code, be fluent in the scientific stack (NumPy, pandas, scikit-learn), and be comfortable structuring a codebase rather than working only in scratch notebooks.
Experience extracting and transforming data from PostgreSQL, files, and free-text records into clean training datasets, with sound handling of missing values, duplicates, and schema drift.
Experience training and evaluating machine learning or deep learning models end to end, using scikit-learn, a gradient-boosting library (XGBoost or LightGBM), and PyTorch.
Sound software-engineering practice: version control (Git), writing testable and reproducible code, sensible use of virtual environments and dependency management, and the ability to debug and profile your own work.
Solid grounding in probability and statistics: comfortable reasoning about distributions, uncertainty, and estimation, not only about predictive accuracy.
Ability to work rigorously with messy, sparse, real-world data and to be honest about what it can and cannot support.
Clear technical communication and the ability to collaborate across technical and commercial teams.
Candidates must be legally permitted to work in Hong Kong.
Desirable (any of the following is a strong plus)
Causal inference, selection-bias correction, or off-policy / counterfactual evaluation.
Factor models, quantitative risk, credit risk, or insurance/actuarial modelling.
Recommender systems, learning-to-rank, or decision-focused ("predict-then-optimize") learning.
Exposure to trade, logistics, supply chain, or financial-services data.
We do not expect any candidate to arrive with all of these. We are looking for a strong programmer with solid statistical fundamentals and the maturity to learn the specific methods the work requires. This suits someone in their first or second research role who wants real mentorship on the methods. Candidates who have worked seriously with uncertainty and sparse data in any domain, actuarial, quantitative finance, or econometrics, are encouraged to apply, even without a recommender-systems background
Compensation and Location
Location: Kowloon Bay, Hong Kong. Full-time, on-site.
Salary: HK$30,000 per month for candidates holding a bachelor's or master's degree, HK$40,000 per month for candidates holding a doctorate.
Pay: $30,000.00 - $40,000.00 per month
Work Location: In person