About
Cheaper, faster, reproducible inference
Persistent agents re-send enormous amounts of stable context. Zumik treats that repetition as a measurable, optimizable quantity - and refuses to dress up provider caching as a product we invented.
Caching is becoming a provider feature. The control plane is the product.
OpenAI, Anthropic, Google, and others now cache prompts natively. So the durable opportunity is not a standalone cache - it is the layer around caching: stable state semantics, reproducible routing, honest QoS accountability, deletion evidence, and the diagnostics to know when any of it is worth doing.
That is what Zumik is. We start by measuring how much repeated work a workload actually has, capture what managed providers already offer, and escalate to BYOK or self-hosted BYOC only when replay data proves the added complexity pays for itself. Long prompts alone never justified self-hosting, and we built the product to say so.
Principles
Non-negotiable design constraints.
Strict at /v1
No proprietary fields on the OpenAI surface. Compatibility is a contract, not a marketing line.
Native at /v2
State, branches, replay, purge, and telemetry live in a surface built for them.
Opaque identities
Public handles are random and tenant-scoped. We never expose raw content hashes.
Logical ≠ physical
A context object can be reusable even when a physical KV realization is not.
Providers first
Managed execution is the default. Self-hosting is earned with replay evidence.
Measure honestly
Opportunity and capture are reported separately, each with an evidence level.
Security as baseline
Tenant isolation, secret hygiene, and purge evidence are requirements, not add-ons.
Replay is a product
A versioned protocol and signed reports, not a one-off script.
We build on the best open infrastructure.
Bifrost for the gateway tier; Dynamo, SGLang, FlashInfer, LMCache, and Mooncake for self-hosted KV orchestration; llm-d, KServe, and AIBrix for Kubernetes. Zumik is the control plane on top, not a reinvention of the runtime.
Build agents that don't pay twice for the same context.
Start with a diagnostic, migrate with one base-URL change, and let evidence drive every optimization.