ErdosFL is an open foundation for collaborative, privacy-preserving AI. Two products, one principle: many parties build a shared outcome together — and no one hands over their data.
ErdosFL spans the federated stack — from training a shared model across private data, to running governed multi-agent systems. Each product is open source and reads in an afternoon.
A lightweight, extensible runtime for federated learning. Many collaborators train one model — sites send only updates, and the raw data never leaves home.
Explore ErdosFC →An agent lifecycle platform — build, test, deploy, reinforce, and govern AI agents from one observable, audited runtime. Five tightly-integrated layers.
Explore ErdosFAI →Beyond the two products, this is where I work with external open-source frameworks day to day — studying them, contributing back, and folding the best ideas into ErdosFC and ErdosFAI. These aren't my products; they're collaborators' work I build alongside.
A global consortium training a shared frontier model — partners then build and own sovereign models aligned to their own needs. Its PoC consortium loop — a coordinator, sovereign nodes, weight-delta sharing under a governed policy — maps straight onto ErdosFC, seeding ideas like quality-floor / anti-capture aggregation.
An evolving set of open-source projects I learn from and contribute to — the best of each finds its way back into the two products.
Paul Erdős published with more than 500 co-authors — his whole body of work was built on collaboration, immortalized by the Erdős number. Both ErdosFL products are collaboration of exactly that kind: many parties — sites, or agents — work toward a shared outcome while their private data stays home. ErdosFL makes that pattern small enough to read and easy to extend.
"Many minds, one result — and nobody hands over their notebook."
Whether the collaborators are hospitals training a model or agents running a governed pipeline, the contract is the same: share outcomes, not raw data.
ErdosFC and ErdosFAI are different runtimes, but they share a spine.
Data and raw weights stay local; only updates or outcomes cross the wire.
Small, swappable abstract base classes — bring your own aggregator, model, or agent.
Dependency-light cores you can read in an afternoon and use as a research scaffold.
Governed, observable runtimes — built to be inspected, traced, and trusted.
Pick a product, clone the repo, and run it offline in one command — no keys, no data leaving home.