Show the model, the dataset, the metric, and the business outcome. ATS-clean, recruiter-scannable, and built for the hiring funnel that screens on PyTorch, SQL, and a deployed model in production.
Recruiters can't tell a serious data scientist from a Kaggle hobbyist from 8 seconds of CV scanning. Help them: name the deployment context (real-time inference behind a gRPC service, batch scoring in Airflow), the training data scale, and the business metric the model moved.
Skills should split into 4 groups: ML/DL stack (PyTorch, sklearn, XGBoost), data engineering (SQL, Spark, dbt, Snowflake), production (FastAPI, MLflow, Sagemaker), and visualisation (Tableau, Looker, matplotlib). Don't dump everything into one block.
Publications and Kaggle ranks are worth one line each at the bottom, never the lede. The lede is the deployed work.
Help: if you're early-career or transitioning from research. A clean Projects block with 3 deployed models, each with dataset, technique, and outcome, beats a sparse work history.
Hurt: if you're 5+ years in industry. A Projects section then implies you don't have enough work experience to fill the page, which is the opposite of what you want.
If you do include projects, treat each one like a job: company-equivalent line, dates, 2-3 quantified bullets.
Reverse-chronological work history with deployed model outcomes, a grouped Skills section (ML, data engineering, production, visualisation), Education with degree and thesis topic, and Publications or Kaggle results only if material.
No. List frameworks (PyTorch, sklearn, XGBoost, Hugging Face), tools you'd build production systems with (MLflow, Sagemaker, Ray), and ecosystem (pandas, numpy). Skip every individual viz library and every minor experiment-tracking tool.
A grandmaster rank, yes, in a one-liner near the bottom. A few finished tutorials, no. Recruiters discount Kaggle as a primary signal because the data is too clean and the metric too narrow.
Name the dataset size, the technique, the baseline, and the lift. 'Trained an XGBoost classifier on 14M labelled transactions; lifted fraud recall from 71% to 84% at the same false-positive budget' is defensible. 'Built state-of-the-art model' is not.
Yes, prominently. Most data science JDs screen on SQL as a hard required skill and 30% of applicants leave it off because they assume it's implicit. Don't leave it off.
Free. PDF or DOCX. No credit card.
Check my CV →