hey flower team - been following your work on federated learning and thinking about a gap in the stack.
flower solves private compute - nodes train locally, share updates without sharing raw data. but the coordination layer (node-to-aggregator, node-to-node) still has trust assumptions. whoever runs the aggregator sees metadata: which nodes are participating, when they’re active, update timing patterns.
for cross-org federated learning (hospitals, banks, competing companies) this metadata can be as sensitive as the data itself.
i lead agents at xmtp - we’re the messaging layer for ai agents. e2ee using mls protocol, decentralized, no intermediary can read or see who’s talking to whom. live in coinbase wallet and world app as infra.
potential fit with flower:
-
node coordination - e2ee transport for gradient/model updates
-
aggregator comms - private signaling for round starts, completions
-
cross-org fl - removes the “who do we trust to run the server” problem
-
agentic fl - as nodes become more autonomous, they need private coordination
just think there’s an interesting intersection here. happy to chat if this resonates.
sdk: https://github.com/xmtp/xmtp-js/tree/main/sdks/agent-sdk
docs: https://xmtp.org/