Skip to main content
Peninsula School District · Gig Harbor, WashingtonOpen source · free to run
Peninsula AI
ProductionMIT licenseUpdated Sep 25, 2026

AI Studio

Frontier AI models for every staff member and student in the district, behind district sign-in, at about a tenth of the cost of per-seat licences.

aistudio.psd401.ai
The AI Studio home screen, with the district name above a short description of the platform

AI Studio gives staff and students access to frontier models — GPT-5, Claude Opus, Google Gemini — through the district's own sign-in, instead of a separate per-seat licence for each one.

We built it because the per-seat maths did not work. Individual AI subscriptions run $20 to $200 a month per person. Calling the same models through their APIs, behind our own authentication, costs about a tenth of that, and we decide where the data goes.

What it does

  • Chat with several models in one place. Switch between GPT-5, Claude Opus and Gemini mid-conversation, or run the same prompt against two side by side and compare.
  • Build an assistant without code. Assistant Architect lets a teacher chain prompts together, pass results from one step to the next, and attach a set of documents — then share the result with colleagues who never need to learn prompting.
  • Ground answers in your own documents. Upload PDFs, Word files and text, OCR included, and assistants answer from them.
  • Keep control at the district level. Role-based access, tool-level permissions and an audit log.

What did not work

We started with Amazon Bedrock Guardrails set to block unsafe content. It kept flagging legitimate educational content. After repeated false positives, we switched it to detect-and-log: detections are recorded for review, and nothing is blocked. That trade-off is deliberate, and it is documented in the repository.

Running it in your district

The code is MIT-licensed and deploys to your own AWS account; the deployment guide is in the repository. If you would rather talk it through first, email us.

Technical specification

Models
GPT-5, Claude Opus and Google Gemini, switchable per conversation
Front end
Next.js 16, TypeScript
Hosting
AWS — ECS on Fargate, with RDS
Sign-in
AWS Cognito with Google sign-in
Model data
Covered by the model providers' zero-data-retention agreements
Content safety
Amazon Bedrock Guardrails in detect-and-log mode
Licence
MIT

Run it in your district

The code is in the repository. If you want to talk it through before you start, email us.