Engineering Notebook

Bring your own deployment

Use your provider account without giving this site your key.

Clone the same tested application into your Vercel account, store credentials as server-side environment variables, and keep billing and operational control inside your deployment.

0keys shared with this site
3allowlisted live samples
300maximum output tokens
1shared daily ceiling

Setup

Deployment checklist

  1. Clone into Vercel. Keep the repository root as the project root; the checked-in configuration serves docs and deploys the Python functions.
  2. Create a dedicated OpenAI project key. Put it only in Vercel as OPENAI_API_KEY.
  3. Install Upstash Redis from the Vercel Marketplace. Link it to the project so the REST URL and token are injected.
  4. Set the remaining variables. Start with live mode disabled, verify health, and then enable it deliberately.
  5. Redeploy and verify. Check /api/health, run one allowlisted sample, and inspect the daily-budget telemetry.

Server-side configuration

Required environment variables

NameExamplePurpose
OPENAI_API_KEYSecret project keyProvider credential; never exposed to the browser.
OPENAI_MODELgpt-5.4-mini-2026-03-17Pinned model behavior and pricing envelope.
TRIAGE_DEMO_ACCESS_CODEStrong private codeRequired gate for every live request.
TRIAGE_DAILY_LIMIT100Maximum globally claimed live requests per UTC day.
TRIAGE_RATE_LIMIT8Per-instance, per-client requests per minute.
TRIAGE_ALLOWED_ORIGINCanonical Vercel URLAllowed browser origin; set after the first deployment.
TRIAGE_LIVE_ENABLEDfalse then trueEmergency live-call kill switch.
UPSTASH_REDIS_REST_URLMarketplace injectedShared counter endpoint.
UPSTASH_REDIS_REST_TOKENMarketplace injectedShared counter credential.

Security boundary

Do not paste a key here

“Deploy your own” means the credential goes directly into your hosting account. This public site never asks for, receives, proxies, or stores a visitor's provider key.

Extension point

Add providers behind one contract

The Python service depends on a narrow provider interface. A later adapter can target another structured-output service while preserving the schema validator, sample policy, telemetry, spend guard, and regression cases.