Open Source Startup Podcast E186: Unlocking Your Unstructured Data with Typedef
12 snips
Nov 20, 2025 Kostas Pardalis and Yoni Michael, co-founders of Typedef, discuss their journey from major data infrastructure firms to creating solutions for unstructured data. They reveal how traditional systems like Spark fail under current AI demands and introduce Fennec, a tailored DataFrame library for LLMs. Insights on optimizing agentic systems and early go-to-market strategies add depth to their conversation. They also ponder the future of AI in data workflows and critique the explosion of benchmarks in the industry.
AI Snips
Chapters
Transcript
Episode notes
Remove Ops Burden Early
- Abstract operational burden so engineers focus on business logic instead of cluster sizing or partitioning.
- Build serverless-style ergonomics to remove ops friction from data teams' workflows.
Make Inference Planner-Aware
- Inference changes workload characteristics from CPU-bound to I/O-heavy and black-box UDFs break optimizer value.
- Treat inference as part of the planner so semantic operators and optimizers can reason about models.
DataFrame API For LLM Workflows
- Fennec exposes a DataFrame API extended with semantic operators like semantic filter and semantic join.
- It batches inference, handles rate limits, and unifies relational and LLM transforms ergonomically.