products · ai
Everything between your model and your product.
Calling a model is easy; getting from a prompt in a notebook to a workload in production is not. The AI tier closes that gap — routing, execution, orchestration, and retrieval, designed as one system.
gateway
One door for every model.
Point your code at one endpoint and call models by alias. Behind each alias is an ordered chain of upstreams, tried in fallback order — so an upstream outage becomes a routing decision, not a war room.
Connect the model endpoints you already use; their credentials live as secret references, never as plaintext in config.
Code calls a model by name; you decide which chain of upstreams serves it. Change the chain, not the code.
Each model shows whether it is genuinely serving — verified from the data path, not assumed from a config entry.
Cap requests or tokens per key, per window. The limit is enforced where the call lands, not in a dashboard nobody watches.
Every completion is attributed to the key that made it, and rollups break usage down by key and by model.
Send a prompt through the gateway and see the response, without leaving the console.
for the finance team
Model usage stops being a mystery and becomes an attributed, capped metric — per key, per model, per window.
sandboxes
Built to vanish.
Some work should never outlive itself. Sandboxes launch untrusted code, evaluation runs, and batch fragments with an optional time-to-live — and are destroyed when they're done.
When the clock runs out, so does the sandbox.
A first-class action, not a cleanup script.
A sandbox receives exactly the secrets and permissions it was granted — nothing ambient, nothing inherited.
Input, output, and status on the same timeline as everything else the platform runs.
Real compute when it matters, gone when it doesn't — one execution story, not two.
workflows
Durable orchestration over the whole platform.
Workflows chain the platform's own capabilities into versioned step graphs: call a model through the gateway, run a command in a sandbox, query an index, wait, branch, repeat.
Updating a workflow publishes a new version — running history stays honest.
Each step is checkpointed as the workflow runs, so a long pipeline survives interruption and picks up where it stopped.
Runs appear on a live timeline with their status; cancel the ones that shouldn't finish.
Kick off a run with JSON input and watch it on the run timeline.
for the engineering team
The nightly index rebuild, the multi-step data prep, the model-eval sweep — they become reviewable, rerunnable graphs instead of fragile cron folklore.
knowledge graphs
Your data, as relations.
Retrieval shouldn't require a separate stack. Knowledge Graphs turn the data you already hold into entities and relations you can query — with the platform taking no custody of it.
Point at the data you already hold; nothing moves, and nothing is copied away.
Model the domain as nodes and typed edges — traversal is the query.
A staged, declarative build from raw data to a queryable graph — resumable, repeatable, visible.
Walk neighbors from any node right in the console before you trust it in a run.
Retrieval is a step in the graph, not a side quest of scripts.
for the data team
The retrieval substrate is a product surface — declarations, pipelines, and query — not a pile of ad-hoc scripts.
vector indexes
Similarity, as an index.
Semantic lookup over your content, treated like infrastructure: declared, versioned, and rebuilt by workflows — never regenerated by whoever still has the script.
An index is a named resource in your project — typed, inspectable, and owned.
Blend nearest-neighbor search with structured filters, so answers stay in the right lane.
Index builds are workflows — versioned, resumable, and visible on the run timeline.
An index reports whether it is genuinely serving — verified, not assumed.
for the platform team
Search stops being a separate stack and becomes another accountable resource in the project.
contact
Talk to us about running Quark.
Tell us about your workloads — we will show you what your own cloud can do.