Sarah Sachs leads AI modeling at Notion, shaping the search, evaluation, model infrastructure, and agent systems behind its workplace AI products. Her work addresses a practical challenge facing every applied-AI company: making increasingly autonomous software reliable, affordable, and safe around sensitive organizational information.
Sachs studied applied mathematics and computer science at Brown University and began her career at Google, working on Google Maps and personalized recommendations. She became a founding machine-learning engineer at Sunshine, formerly Lumi Labs, and later led natural-language processing and generative AI at Robinhood, where her work included a compliant AI assistant and automated content moderation. At Tome, she served as director of engineering for AI and infrastructure, overseeing AI-generated presentations, model infrastructure, and the company’s OpenAI relationship.
At Notion, Sachs has helped extend workplace AI from writing assistance and database automation into enterprise search, research, and governed AI teammates. In 2026, she announced Notion’s acquisition of search company ZeroEntropy, reporting that its technology accelerated unified search by up to 30 percent and reduced reranking latency by 85 percent. ZeroEntropy founder Ghita Houir Alami joined Notion to lead a new model-research team.
- Product-specific AI evaluation: Sachs designs evaluations around realistic workplace tasks instead of relying on generic benchmarks. Her approach uses carefully curated examples, multilingual checks, user feedback, and example-specific LLM judges that test concrete requirements such as language, formatting, citations, and tool selection. Assessing retrieval separately from generation helps isolate failures when enterprise documents and permissions change. Her AI engineering workshop also emphasizes inspecting individual failures instead of trusting aggregate scores.
- Task-level model economics: Advertised token prices can conceal the real expense of verbose outputs, retries, and slower execution. Sachs evaluates cost per capability per second across complete workflows, reserving advanced models for difficult reasoning while routing routine work to cheaper alternatives. Deterministic operations, including file conversion and structured queries, often belong in conventional software instead of language models.
- Model-agnostic AI architecture: Sachs treats access to multiple model providers as both technical resilience and negotiating leverage. Her Token Town keynote argues that dependence on one supplier exposes companies to outages, forced upgrades, unfavorable pricing, and shifting model capabilities. She has also described testing open-weight models against realistic user tasks to preserve customer choice and reduce the cost of recurring knowledge work.
- Shared context with enforceable boundaries: Sachs sees Notion as a durable system of record where people and agents coordinate around shared documents, permissions, and organizational knowledge. She has warned that fragmented data undermines agents, while emphasizing that autonomous systems handling private information, untrusted content, and external communication need governance, visibility, and protection against prompt injection. Integrations involving Claude, Codex, and Decagon demonstrate interoperability, not ownership of those external systems.
In 2026, Fast Company included Sachs in its AI 20 recognition, highlighting her work on accessible, governed AI agents.
