Where to apply ML in a task flow from customer request to driver-uploaded evidence

This was a short, free-form design discussion at the end of a phone screen. The interviewer described a platform on which customers submit tasks that drivers on the platform carry out: a customer places a task, a driver completes it, and the driver uploads evidence that the task was done. The question was: **Across this end-to-end flow, from the moment a customer submits a task until a driver has completed it and uploaded evidence, where could machine learning, including large language models, be leveraged?** Map the flow, identify the opportunities, and go deeper on the two or three you consider most valuable: what each model would do, what data and labels it would learn from, how you would evaluate it, and what could go wrong. ```hint Find the human judgment calls Walk a task from the customer's request to the uploaded evidence and mark every step where a person currently reads, writes, decides or checks something by hand. ``` ```hint Labels decide feasibility For each candidate opportunity, ask what a wrong prediction would cost at that step and whether past outcomes already provide labels to train and evaluate a model. ``` ### Constraints and Clarifications - The task types, evidence formats, volumes and current tooling were not specified. Ask about them or state your assumptions. - The discussion is short: cover the whole flow first, then go deep on a few ranked opportunities rather than every idea. ### Clarifying Questions - What kinds of tasks do customers submit, and what counts as evidence: photos, videos, documents, text, location data? - How do customers describe a task today: free text, a form, or a conversation with a sales or operations team? - How is evidence reviewed today, and how many submissions arrive per day? - Which error is costlier: accepting bad evidence, or rejecting a driver's valid work? - Does a rejection of evidence affect the driver's pay for the task? ### What a Strong Answer Covers - A map of the flow from task submission to evidence upload, with each decision point and hand-off named - Opportunities ranked by impact, feasibility and label availability rather than listed - For the top opportunities: the prediction or generation target, the inputs, the labels, and the choice between an LLM and a task-specific model - Offline and online evaluation, with a human-review path where errors are costly - Risks specific to this flow: fraudulent or recycled evidence, fairness and appeals for drivers, privacy of uploaded content, and the cost and latency of model calls ### Follow-up Questions - At launch you have few labeled rejections of evidence. How do you train and evaluate an evidence-quality model? - A driver disputes an automatic rejection. How does your design handle the appeal, and how do appeals feed back into the model? - How would you detect drivers who upload recycled, staged or edited evidence? - Instructions generated by an LLM turn out to be wrong for some new task types. How do you catch that before drivers act on them?