What the work involves

You write and review machine learning code the way a working practitioner would, and the artifacts you produce — the model, the notebook, the failure analysis, the write-up — are used to teach or evaluate AI systems. A typical task might be building a small end-to-end pipeline against a supplied dataset (preprocessing, feature engineering, training, hyperparameter tuning, benchmark reporting), or critiquing a model's attempt at the same problem and explaining precisely where its reasoning or code breaks. MongoDB features prominently in the customer project: expect tasks involving aggregation pipelines, indexing choices, and pulling training data out of document stores rather than tidy CSVs.

A large share of the value sits in documentation. Reviewers care whether your experiment log is reproducible, whether you stated the metric before you optimized it, and whether you noticed leakage, class imbalance, or a train/test split that quietly invalidated the result. Vague conclusions ("performance improved") are the most common reason submissions get sent back.

What the screen looks for

  • Specific, verifiable framework experience — which of scikit-learn, TensorFlow, or PyTorch you actually shipped with, and on what
  • Whether you can name and defend the evaluation metric appropriate to a given problem, under follow-up
  • Real MongoDB usage rather than passing familiarity: queries, aggregations, schema decisions
  • Whether you've operationalized a model in a cloud or enterprise environment, not only trained one in a notebook
  • Written clarity, since much of the deliverable is prose about code

Logistics

Fully remote, contractor engagement, largely asynchronous with occasional syncs across cross-functional contributors. Volume is project-driven and can fluctuate; most contributors treat it as part-time alongside other work, with a practical floor of roughly 10–15 hours a week to stay in rotation. Pay is hourly and observed in the stated band — micro1 sets rates per project and per contributor level, so treat the range as observed rather than promised.