Flowco: Mixed-Initiative Authoring of Reliable End-to-End Data Analyses via Dataflow Graphs and LLMs
summary
Flowco is a web-deployed system (goflow.co) that composes LLM-generated Python into explicit dataflow graphs whose nodes have a three-level node abstraction (natural-language intent, generated requirements/assertions, executable code). The system iteratively regenerates code until all assertions pass, runs nodes independently, and reproduces real published analyses (sociology multiverse and ML workflows). A user study reports novices quickly learn the canvas and chat integration and keep unchanged parts stable while editing specific steps.
AI-written summary · about
chapters (8)
- 0:00Introduction to FloCo
- 0:44Computational notebooks challenges
- 2:10Need for new programming model
- 2:52Data flow graphs as core abstraction
- 3:56Example analysis in FloCo
- 5:20Design features of FloCo
- 7:48Real-world system and future work
- 8:30Q&A session
related talks
01h 09mFloki: Defining Agentic Workflows for Security
34mOrchestrating Generative AI Workflows to Deliver Business Value
24mLLM App Development Using PromptFlow and Azure Cosmos DB
27mThe Lean LLM development playbook
30mFrom COBOL to Copilots: Why “No-Code” Never Kills the Coder
35mFloki: Building Agentic Workflows with Dapr
40mCode-First LLMOps from prototype to production with GenAI tools
01h 42mAI Powered Application Modernization using Claude Flow