Open source
Three deep projects, one per language. Each ships with a design doc, reproducible benchmarks and a series of write-ups.
Per-task cost, latency and quality telemetry for LLM agents built on Google ADK, with model routing learned from evals and a semantic cache that measures its own false hits. Benchmarked on public agent benchmarks.
LLM evaluationmodel routinginference cachingOpenTelemetry
A small durable-execution engine for long-running agent workflows: Raft-replicated log, leases, idempotent activities, SLO-based alerting and fault-injection tests with published postmortems.
consensusexactly-once semanticsSREfault injection
A columnar time-series store with Gorilla and delta-of-delta compression and vectorized execution, exposed as a DuckDB extension that supports federated queries with predicate and aggregate pushdown.
storage formatscompressionSIMDquery pushdown