Lead Data Scientist (Agentic Solutions)

Rivian•Irvine, CA

77d•Onsite

About The Position

Rivian's AI team sits within the Analytics function of the Customer organization—the team that spans Marketing, Sales, and Market Intelligence—and is building the next generation of intelligent, autonomous systems that connect these functions into a single, reasoning ecosystem. As a Staff/Lead Data Scientist focused on Agentic Solutions, you will design and operationalize the cognitive architecture that powers Rivian's AI agents—building the reasoning loops, retrieval systems, and evaluation frameworks that allow agents to act on live business data with accuracy and reliability. This is a deeply technical, high-impact role at the frontier of applied AI, working closely with Context Engineering, Data Platform, and the broader Customer org Analytics team to bring agentic intelligence to production.

Requirements

Bachelor's degree in Computer Science, Statistics, Mathematics, Engineering, or a related quantitative field; Master's or PhD preferred.
5+ years in Machine Learning, Applied Data Science, or AI Engineering, with a proven track record of designing cognitive frameworks, RAG systems, or agentic workflows for business applications.
Expert-level capability with modern AI orchestration frameworks (LangChain, LlamaIndex, Haystack) and a deep understanding of LLM APIs, prompt engineering methodologies, and agent memory state management.
Hands-on experience working with SLMs (e.g., Phi, Mistral, Gemma) for domain-specific deployment; ability to reason about the trade-offs between model size, latency, cost, and task accuracy. Experience with parameter-efficient fine-tuning (LoRA/QLoRA) and model distillation.
Deep understanding of token economics; experience designing agent frameworks with context compression, prompt caching, dynamic windowing, and call-minimization strategies to build systems that are both intelligent and economically viable at scale.
Expert-level Python and highly proficient SQL. Comfortable inside Databricks (Spark, Delta Lake) navigating structured relational schemas, metadata, and vector indices to extract and optimize model context.
Strong statistical foundation with familiarity analyzing user behavioral data or event-streams (e.g., Snowplow) to help agents deduce funnel health and operational anomalies.
Proven ability to operate in a fast-paced, high-ambiguity environment—such as a new product launch or startup-stage AI build—with intellectual curiosity, rigorous attention to detail, and a bias for shipping.
Ability to stand, sit, or walk for 8-10 hours per day.
Required to communicate using phone and/or e-mail.
Ability to view, read, and interpret documents.
Ability to perform all duties in an office environment that may contain ambient noise and temperature fluctuations.

Responsibilities

Design and program multi-step reasoning frameworks using orchestration frameworks like LangChain, LlamaIndex, or Haystack.
Architect advanced retrieval structures, experimenting with embedding models, dynamic chunking strategies, and token management to ensure agents receive high-fidelity business context.
Build the internal logic and evaluation criteria for 'Watchdog' agents, enabling them to analyze live operational data streams, infer anomalies, and formulate proactive insights.
Evaluate, select, and adapt small language models (e.g., Phi, Mistral, Gemma) for domain-specific agentic tasks where precision, latency, and cost efficiency outweigh raw model scale. Design the decision framework for when to use an SLM vs. a frontier model based on task complexity, context requirements, and inference cost.
Architect agents and orchestration frameworks around token efficiency as a first-class design constraint—developing strategies for context compression, prompt caching, dynamic context windowing, and call minimization to ensure every inference call is purposeful and cost-effective at scale.
Execute parameter-efficient fine-tuning (e.g., LoRA/QLoRA) and model distillation on open-source models to embed Rivian's fulfillment jargon, vehicle logistics vocabulary, and internal business rules—prioritizing targeted SLM adaptation over large-scale LLM retraining where feasible.
Establish statistical, model-driven, and human-in-the-loop testing benchmarks to empirically validate agent reasoning, track accuracy drift, and minimize hallucinations.
Continuously iterate on complex system prompting, structured output formats (e.g., enforcing strict JSON/tool schemas), and cognitive guardrails to keep agent behavior deterministic and aligned.