When 97% of companies say they’re struggling to show value from their generative AI initiatives, something is clearly broken. Enterprises are pouring resources into pilots, prototypes, and proofs of concept, only to watch them stall before ever reaching production. It’s a phenomenon the industry now calls pilot paralysis.
Into this paralysis steps Typedef, a new AI infrastructure startup that officially launched on June 18, 2025, with $5.5 million in seed funding. Led by Pear VC and supported by Verissimo Ventures, Monochrome Ventures, Tokyo Black, and several angels, Typedef’s goal is to make AI work at scale.
Their promise? To turn brittle, non-deterministic LLM prototypes into reliable, production-ready workloads, at speed and with the rigor of traditional data pipelines.
A Crisis of Scale in AI
The AI infrastructure market is projected to reach $200 billion by 2028. Yet despite the excitement, most AI initiatives fail to generate business value. That’s not because the models aren’t good.
“Legacy data platforms weren’t built to handle LLMs, inference, or unstructured data,” said Typedef co-founder Yoni Michael. “As a result, the workaround has been a patchwork of systems, aging technologies and tooling, or DIY frameworks and data pipelines that are brittle and unreliable”.
It’s a problem Michael and his co-founder, Kostas Pardalis, have felt firsthand. Both are data infrastructure veterans, with careers that span Salesforce, Starburst, and Tecton. They’ve seen firsthand the midnight headaches of debugging Spark jobs, scaling clusters manually, and watching complex inference workflows crumble under real-world loads.
Typedef: The AI-Native Infrastructure Engine
Their solution is a new AI data engine built from scratch for inference-first, LLM-driven workloads. Typedef is fully serverless: no infrastructure to provision, no clusters to configure. Users simply import its open-source SDK, connect their data, and begin building production pipelines using familiar abstractions like Python and DataFrames.
The secret lies in its architectural philosophy: inference is a core data transformation.
“Typedef gives you a unified engine for building AI-native data pipelines and agentic applications that span structured tables, messy text, embeddings, and LLM outputs,” Michael wrote in a blog post.
What they say sets Typedef apart is its ability to tame AI complexity. It manages token limits, context windows, chunking, rate limits, retries, everything that makes production LLM pipelines a nightmare. And it does it with a clean, composable API.
“It’s about delivering deterministic workloads on top of non-deterministic LLMs,” Pardalis said. “AI and data teams want the same rigor and reliability they expect from traditional data pipelines”.
Early customer results suggest Typedef’s approach is working. Take Matic, an insurance technology platform that partners with over 70 carriers. According to Chief Product Officer Lee Maliniak, deploying Typedef let them “build and deploy semantic extraction pipelines across thousands of policies and transcripts in days, not months.” The result: reduced human error, significantly lowered costs, and decreased Errors and Omissions (E&O) risk.
Fenic: Open-Sourcing the Future of AI Workflows
Typedef has open-sourced a key part of its engine: Fenic, a PySpark-inspired DataFrame framework purpose-built for AI workflows. Fenic handles semantic operators, batch inference, and unstructured data types (like markdown and transcripts) while preserving columnar consistency, auditability, and lineage.
Why open source such a core component?
“Open sourcing Fenic expands the ecosystem,” Michael and Pardalis wrote. “We stand on the shoulders of projects like Apache Arrow and DuckDB. We want to build with the community and make AI systems more production-ready, together”.
Fenic allows developers to preprocess, enrich, and transform unstructured data for downstream AI agent, all while maintaining the composability and determinism enterprise teams require.
Typedef’s Roadmap
With fresh funding and early traction, Typedef is looking ahead. The company plans to expand support for more semantic operators, agentic pipelines, and data sources in the months ahead. Its cloud engine is currently in alpha, with a full release expected later this year.
“We’re just getting started,” the founders wrote. “If you’re building production AI systems-or want to-we’d love to hear from you. Let’s make AI actually work”.
