Senior Machine Learning Engineer

Core Responsibilities:

.Be the key owner for modeling research execution for client classification and related risk/anti-fraud checks frame hypotheses with Quants, align labels with RA, and run clean, time-aware evaluations.

.Broaden scope to risk/anti-fraud: apply the above methods to sibling controls along the transaction chain (toxic-flow & scalping patterns, slippage anomalies, stop-out risk, collusion/copy-trading, funding-rate/credit abuse, withdrawal risk) prioritize by expected P&L and operational impact.

.Own the experiment loop end-to-end (hypothesis backtest thresholding/calibration monitored rollout) with MLflow tracking and clear readouts.

.Optimize for business impact: tune decisions for net P&L, and FP$/FN$ make trade-offs explicit (accuracy vs. latency/cost).

.Ship safely to production: shadow tests, canary releases, rollback plans, and lightweight guardrails aligned with risk policy.

.Keep models healthy: monitor drift, label delays, and calibration set retrain triggers and maintain a steady weekly iteration cadence with Quants/RA.

.Communicate clearly: short design docs, experiment summaries, and release notes that non-ML stakeholders can act on.

In Addition:

.Time-series modeling of order streams (when it helps): comfortable prototyping and comparing simple sequence models (e.g., an LSTM or a light Transformer) against straightforward baselines pick the simplest approach that moves the metric.

.Practical representation learning: able to learn compact client/strategy embeddings to improve cold-start and data efficiency (e.g., a basic autoencoder or simple contrastive objective) without overcomplicating the stack.

.Working with imperfect labels: experience basic approaches for partial/noisy labels-e.g., positive-unlabeled setups or simple denoising checks, so models stay robust.

.Policy impact & efficient iteration: estimate offline impact of routing decisions with lightweight methods run small, well-instrumented HPO jobs (e.g., Optuna/Hyperopt) and keep costs/latency in check.

.Uncertainty & deferral: provide calibrated scores with simple confidence/deferral logic when unsure, safely fall back to rules or RA review.

Qualifications:

.5+ years owning ML systems end-to-end with a track record of modeling research production impact on large tabular/time-series problems.

.Strong PySpark/Spark SQL on Databricks and Python excellent command of MLflow (or equivalent), model registries, CI/CD, and jobs-as-code.

.Proven skill in experimental design: temporal cross-validation, leakage prevention, out-of-time back-tests, label-delay handling, and calibration/thresholding.

.Comfortable optimizing for asymmetric costs and communicating trade-offs (precision/recall vs. P&L, latency vs. accuracy) to non-ML stakeholders.

.Clear, concise writing able to own RFCs and implementation plans.

.Experience with partially labeled or noisy-label settings, active feedback loops, and production A/B or shadow testing.

.Practical interpretability for stakeholder trust and release gating.

.Comfortable to use chinese and english as daily working language to work with chinese speaking stakeholders

Tech you'll use

.Databricks (Spark/Delta/UC/MLflow), Python, Spark SQL, Databricks Workflows/Jobs