Job DescriptionTeach AI how business actually worksAI agents are starting to take on real business work: releasing blocked invoices, re-planning supply, preparing the financial close. To do it safely, they need more than access to data. They need to know what an order, a supplier and a cost center are, how they connect, which rules and authorizations apply, and what happened the last time someone made a similar decision. When executing mission-critical business processes, an answer or action that is right 90% of the time is not good enough.
SAP systems hold the context for many of the world's largest companies. SAP's knowledge graph for SAP S/4HANA alone is based on 452,000 tables, 80,000 CDS views (semantic data models) and 7.3 million fields. Yet much of the meaning, especially in each company's own extensions, configuration and decision history, is still implicit. A 'customer' in sales, a 'business partner' in finance and an 'account' in CRM can be the same company described three ways, and an agent deciding whether to raise a credit limit has to know that. This is not a data-cleaning problem, it is a knowledge-representation problem. Today that context, the semantic layer agents rely on, is mostly modeled by hand, one company at a time. We think it can be learned, so AI agents can act on business data, not just read it. SAP's Autonomous Enterprise depends on that, and it is what makes the agents people work with in Joule Work, SAP's AI workspace, accurate enough to trust.
As a Senior Data Scientist, you'll own a research thread end to end: from an open question, to a method and a benchmark, to a prototype a product team adopts. This is a research and modeling role, not product analytics.
SAP Research & Innovation North AmericaWe are a team of scientists and engineers at SAP Labs in Palo Alto, minutes from Stanford. We look a few years ahead and scout emerging technologies, take the early, riskier bets that product teams can't always take, and prove or drop them fast. SAP's product teams build and run SAP Knowledge Graph, we look beyond and find and prove methods that demonstrate improvement or paradigm shifts. Our job is to spot the opportunities that will matter most to SAP and our customers, and to get there early. We prototype with dedicated GPU compute on realistic business data and, in co-innovation projects, on real data for training and evaluating models. We publish research, work with startups, universities and institutes including Stanford HAI and UC Irvine, and hand what works to SAP's product teams. We move like a small lab, with the reach of a company whose software runs finance, supply chain and HR for many of the world's largest companies.
What you will work on- Research you own. Take an open question from ambiguous framing to a reproducible result, and decide what to do next based on what you find. Closely collaborate with top universities and institutes such as Stanford HAI.
- What comes next. Reproduce promising new papers on SAP data within weeks and recommend which deserve a bigger bet.
- Better ways to build the semantic layer agents need. New methods to learn and maintain ontologies of business objects, processes, rules and actions, with SAP's semantic models as the backbone and LLMs, schema matching and entity resolution to extend them.
- Agents grounded in business context. Graph-augmented retrieval and text-to-query agents (SQL, SPARQL, APIs) over very large schemas, with checks against business rules and access rights, and governed actions.
- Models that learn from relational data. Graph neural networks and relational foundation models, such as SAP-RPT-1.6, applied with knowledge-graph context to prediction, anomaly detection and what-if analysis.
- Honest measurement. Datasets, benchmarks and evaluation harnesses that show whether an approach works on realistic enterprise data, and why, using synthetic or privacy-preserving data where real data can't be shared.
- Work others can use. Clean Python that others build on, prototypes built together with (and adopted by) product and engineering teams, published research papers or open-source releases, and mentoring for interns.
In your first year, success could look like:- Own one research thread end to end, hands on. For example, schema linking for agents over SAP finance data, taken from baseline to benchmark to a prototype a product team adopts.
- A paper submission or open-source release.
This is a senior individual-contributor role, comparable to Senior Research or Data Scientist at other Bay Area technology companies.
QualificationsRequired- Typically 5+ years of hands-on data science, machine learning or applied research, or a PhD plus 2+ years, taking projects from an open question to a measurable result. An MS or PhD in computer science, machine learning, statistics, mathematics or a related field, or equivalent practical experience.
- Hands-on depth in at least one of the following, and an appetite to learn the others:
- Knowledge graphs, ontologies or semantic layers: RDF, OWL, SHACL, SPARQL, or property graphs with Cypher or similar
- Entity resolution, schema matching or data integration
- Machine learning on relational, graph or tabular data, including GNNs
- LLM systems over structured data: text-to-SQL, schema linking and API discovery, GraphRAG, agents, evaluation
- Solid foundations in machine learning and statistics, with disciplined experimentation: strong baselines, ablations, leakage-safe splits, error analysis and honest reporting of what didn't work.
- Strong Python and SQL, and PyTorch or a similar framework; at ease with large, messy, multi-table data.
- Strong problem solving skills, shown by concrete examples: you have turned a vague question into an experiment that gave a clear answer within weeks.
- Strong communication skills, shown by concrete examples: you have written up a negative or mixed result clearly enough that others acted on it, and you can explain your work concisely verbally and on a one pager to a scientist or executive.
Preferred- Publications at venues such as NeurIPS, ICML, ICLR, KDD, ACL/EMNLP, SIGMOD/VLDB, ISWC, WWW, or contributions to open-source projects.
- Graph databases, vector search, or large-scale data processing such as Spark.
- LLM-assisted ontology learning or knowledge-graph construction.
- LLM fine-tuning or post-training, or MCP.
- Process mining, causal inference or forecasting, and familiarity with enterprise processes such as finance, procurement or supply chain.
- SAP experience is a plus, not a requirement.
If you recognize yourself here but do not match every line, apply.
