Edge AI

What it is

Edge AI hosts Zipper’s inference catalog. Call language, embedding, and classification models from Edge Functions without standing up a GPU fleet. Tokens are billed per workspace, and prompts can be redacted before they hit logs.

Cloudflare analogue: Cloudflare Workers AI. SLA 99.9% · p99 420ms. Admins attach models. Operators may change quotas.

• ▸Language models
• ▸Embeddings
• ▸Prompt redaction
• ▸Per-workspace quotas

How it works

Attach zip-fast, zip-large, or embed-v1 to the workspace. Call from Edge Functions without a GPU fleet.

Prompts can be redacted before Watchtower. Identical prompts may be cached.

Use cases

Concrete ways teams use this service on day one.

Set it up in the dashboard

Dashboard → Services → Edge AI. Attach a model and a task.

API

Control-plane: GET|POST /api/v1/ai. Send Authorization: Bearer tz_live_YOUR_TOKEN.

Creates count against the plan quota. A 402 plan_limit means you are at the cap — upgrade or delete an unused resource.

curl -sS -X POST https://tinyzipper.com/api/v1/ai \
  -H "Authorization: Bearer tz_live_YOUR_TOKEN" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{"name":"Support assistant","model":"zip-fast","task":"text"}'

const res = await fetch("https://tinyzipper.com/api/v1/ai", {
  method: "GET",
  headers: {
    Authorization: `Bearer ${process.env.ZIPPER_TOKEN}`,
    "Content-Type": "application/json"
  }
});
const json = await res.json();
if (!res.ok) throw new Error(json.error ?? res.statusText);
console.log(json);

Tips

Cap max tokens. zip-large is slower and for longer completions.

• ▸SOC 2
• ▸Prompt redaction

Runbook

Cap tokens per call. Redact prompts before they hit The Watchtower. Cache identical completions.