Lars Grammel is a software engineer at Vercel and a leading developer of the Vercel AI SDK, an open-source TypeScript toolkit for building AI applications and agents. He created ModelFusion, the orchestration library that joined Vercel in 2024 and helped shape the SDK’s unified approach to working with different models and providers.
From visualization research to AI infrastructure
Grammel studied computer science at RWTH Aachen University and completed a doctorate at the University of Victoria in 2012, researching interfaces that help novices create data visualizations. At data-preparation company Trifacta, he helped develop its product from its early private-beta period, established its Berlin office in 2015, and led its German operation.
He later developed ModelFusion to coordinate language models, image generation, speech, and embeddings in TypeScript applications. His Storyteller project put those capabilities to work generating illustrated, narrated preschool stories from spoken prompts.
ModelFusion became part of AI SDK 3.1 in May 2024. Grammel subsequently helped expand the SDK into structured generation, observability, tool calling, and agent workflows. He coauthored the June 2026 release of AI SDK 7, which added reasoning controls, tool approvals, durable execution, sandbox integrations, real-time voice, and video generation; Vercel identified him as the release’s primary engineer.
- Provider-agnostic AI development: Grammel’s libraries give TypeScript developers a consistent way to combine or switch model providers while retaining access to provider-specific capabilities.
- Streaming structured generation: Storyteller incrementally parses an unfinished story into complete passages, enabling speech synthesis and playback to begin before the full narrative has been generated. That approach makes structured model output useful while it is still arriving.
- Responsive multimodal orchestration: Storyteller runs title generation, illustration, and narration concurrently; streams updates between its React client and Fastify server; and uses embeddings to match characters with distinct voices. Its design treats latency, narrative consistency, and immediate user feedback as interconnected engineering problems.
- Production-ready agent infrastructure: His later SDK work supports practical safeguards and operations for autonomous software, including tool approvals, telemetry, durable workflows, sandboxed execution, and compatibility with external agent harnesses.
