Cross-Referencing Clients with LLMs: How We Shipped QUIC in Teku

Duration: NaN:NaN:NaN

Speaker: Lucas Saldanha

Type: Talk

Expertise: Some familiarity with the topic

Event: Devcon

Date: Nov 2026

Ethereum has many independent implementations of the same specs, in languages no single engineer reads fluently. That's verification ground truth we don't use. QUIC in jvm-libp2p sat half-finished; nobody on the team had time to finish it. We had an LLM check our code against the networking spec, then cross-reference it with go-libp2p and rust-libp2p. That surfaced spec divergences and unwritten conventions the other clients already followed: flow-control windows hardcoded at 1MB where siblings use 15MB, and a stateless-reset token that only interops if its derivation matches theirs. The gap list is jvm-libp2p#476; closing it shipped QUIC in Teku (#10805), now on by default. LLM drafted the test plan, too. The resource-leak investigations took JVM internals and protocol knowledge. The claim is capacity, not just productivity: small teams carry real work that never gets scheduled but still needs to be done, and that bounds client diversity. LLMs can help fill the gap.