Staff Data Scientist (Tech lead)
DevSavant Inc., LATAM
The role
We’re looking for a Staff Data Scientist to join a fast-growing U.S. data and analytics company. You’ll own one of the hardest problems in the business: matching records across large, independent consumer datasets that were never designed to join, and turning incomplete, noisy data into numbers clients trust.
This is a high-autonomy role. You’ll take loosely scoped problems and drive them to validated, production-ready solutions. You’ll also partner closely with Engineering and be the person the team brings its hardest questions to.
What you will do
- Own and evolve the core matching algorithm that probabilistically links two independent datasets, including sample size, match-window design and reproducibility of the production pipeline.
- Own eligibility logic: decide which records qualify for matching and reporting, balancing data quality against reach and sample size.
- Run the recurring matching process end to end, improving match quality and runtime cost over time.
- Investigate and resolve cases where models interact or conflict, including unexpected second-order effects.
- Partner with Engineering on how models are architected, deployed and scaled (dbt, BigQuery, GCP), treating cost and runtime as seriously as accuracy.
- Own the written methodology: design docs, decision records and version history.
- Lead methodology change management: decide when a change requires restating history, version each release, and protect trend continuity for clients.
- Mentor teammates on investigation methods and code quality.
- Present methodology decisions and trade-offs directly to senior leadership.
What you need
- 7–10 years in data science or applied science, with real ownership of production systems, not just research.
- Experience reconciling or matching large, messy, multi-source consumer datasets, such as identity graphs, credit or enrichment data, clean-room data, or fraud and identity work.
- Deep expertise in entity resolution and record linkage: blocking strategies, probabilistic matching (e.g., Fellegi-Sunter), match-confidence scoring and precision/recall trade-offs.
- Strong panel and sampling statistics: weighting and raking, bias correction, and minimum reporting thresholds for small samples.
- Uncertainty quantification with confidence intervals and bootstrapping, and the ability to tell a real period-over-period change from noise.
- Expert, production-grade Python (OOP, tested, maintainable) and strong SQL on BigQuery.