Most data scientist searches fail before the first message is sent. The req says "data scientist", the hiring manager means "someone who can run our A/B tests", the recruiter searches for machine learning keywords, and the shortlist is full of people who would be bored by the job. Fixing the brief is cheap. Fixing a bad hire six months later is not.
Three jobs, one title
Analytics and experimentation. SQL heavy, close to product managers, owns metrics, dashboards and test design. The question they answer is whether a change worked.
Statistical modeling. Forecasting, pricing, risk and causal inference. Often comes from economics, statistics or the sciences, and is comfortable in R as well as Python.
Production machine learning. Builds models that run inside the product, from ranking to fraud detection, and cares about pipelines, latency and monitoring. This person overlaps heavily with machine learning engineers.
Ask the manager which decision this hire will change in the first quarter. The answer tells you which bucket you are in.
What the survey data says about the toolset
In the Stack Overflow 2025 Developer Survey, 57.9% of all respondents reported using Python, after a 7 point jump from 2024, while R was at 4.9% and Jupyter notebooks at 14.1% among development environments. Two practical consequences follow. Python on a profile is close to universal and tells you little; R is rarer and is a useful signal for statistical roles. And because notebooks are a minority tool among developers overall, a public notebook habit is a good marker of someone who actually does analysis.
Where the work is visible
GitHub. Look for repositories with real data handling, tests and a readable README rather than course assignments. Contributions to libraries such as scikit-learn, pandas or PyTorch are strong signals.
Kaggle. Competition rankings and, more usefully, the write ups that explain an approach. Medal winners in tabular competitions are often excellent applied modelers.
Hugging Face. Model cards, datasets and Spaces show who is training and evaluating models in public, which matters for production machine learning roles.
arXiv and conference proceedings. For research leaning roles, authors on applied papers in your domain are a precise list to start from.
Searching these by hand is slow. Tools that read the open web can help: AI sourcing for data science roles such as Noon evaluates candidates on their technical depth and problem domain, then reaches out by email and SMS.
Outreach that a data scientist will open
Lead with the data and the decision. "We have eight years of transaction data and want to cut false positives in fraud review" beats any line about a fast growing team. Name the stack honestly, say whether models reach production, and say who they would report to. Data scientists who have been stuck building dashboards nobody reads will respond to a role where their work changes something.
A screening loop that tests judgment
Skip the whiteboard algorithm round. A better loop: a 30 minute conversation about one past project, focused on what went wrong and how they knew; a small take home or live exercise on a messy, realistic dataset, capped at two or three hours; and a final discussion where they explain results to a non technical manager. For analytics roles, ask them to design an experiment and name the ways it could mislead you. For production roles, ask how they would monitor a model after launch.
Closing the offer
Strong candidates compare data access and infrastructure as much as pay. Be ready to say what warehouse and compute they will have, how long it takes to get production access, and whether there is a data engineer to work with. If the honest answer is that they will be the first data hire, say so and describe the support they will get.