One endpoint in front of OpenAI, Anthropic, and self-hosted models, with semantic caching, cost- and complexity-aware routing, automatic failover, per-key budgets, and full observability. Why Relay Teams wire the OpenAI SDK straight into production and inherit every operational gap that comes with it: no shared cache, no cost controls, no failover when a provider has a bad five minutes, and no single place to see spend or latency. Relay is a drop-in OpenAI-compatible endpoint that closes those gaps without changing a line of application code, just point your SDK's base_url at it.

It helps you fine-tune, serve, and monitor a document-understanding model in production. Drop in a receipt, get structured JSON with per-field confidence in under 150 ms. It is a live, inspectable system: Fine-tune: LayoutLMv3 token-classification with clean training + evaluation. Serve: FastAPI inference, ONNX Runtime optimization, Prometheus metrics, and enforced p99 latency SLOs. Deployed on Kubernetes via KServe, Triton with HPA autoscaling. Review: a human-in-the-loop UI surfaces low-confidence fields Monitor: input-distribution drift (PSI), accuracy over time, Grafana dashboards.

Chronos is the durable-execution backbone that long-running jobs and agent systems need. You write ordinary Go functions; Chronos makes them crash-proof. Every workflow is event-sourced, so a process can die at any instant and another worker will deterministically replay its history and continue from exactly where it left off, never re-running work that already happened. Deterministic replay of a 1,000-event workflow history in ~61 µs (p99 < 0.15 ms), zero external calls. Kill a worker mid-flight and the workflow resumes on another with zero duplicated side effects.

Give it a goal. It proceeds with using planner–executor–critic; It plans the work, executes it with standards-based tools, verifies its own output, and returns a citation-backed report — with human approval gates, live-streamed reasoning, and durable execution that survives a crash.

BlazeKV matches single-node Redis 7.4 on GET/SET in a single-threaded head-to-head on identical hardware (Linux, io_uring), while adding shared-nothing sharding, a HyperLogLog and HNSW vector module, and Prometheus-native observability, built from scratch in C++20.

BlazeKV is thread-per-core and shared-nothing, in the spirit of Seastar / Dragonfly:

The kernel spreads client connections across cores with SO_REUSEPORT — no user-space accept lock. Each shard owns a hash-partition of the keyspace and runs its own event loop. Single-key commands execute inline with no locks.