Minimal
Public information or information intended for public access. Protecting its integrity is the primary concern.
Hosting and ownership
What we check before a service ships
TritonAI gives campus services approved routes to AI models and campus information. Anyone can review the privacy, security, and accessibility decisions behind those routes on this site.
The platform does not make a service trustworthy on its own. Each service still needs a stated purpose, data it is allowed to use, a named owner, and people who check the results.
How a supported service fits together
- 01Campus serviceA named owner says who it is for, what it does, and how results get reviewed.
- 02TritonAI LLM GatewayOne managed route from supported applications to approved models.
- 03Approved model routes
- Enterprise cloud
AWS, Microsoft Azure, and Google Cloud Vertex AI - UC-hosted
Open models on UC San Diego infrastructure at the San Diego Supercomputer Center
- Enterprise cloud
- 04Approved information sourcesAssistants look things up when someone asks, from sources they are cleared for.
- 05Bounded tools and reviewAgentic services use approved capabilities within a stated scope, with people reviewing consequential results and actions.
UC data classification
What the Protection Levels mean
UC assigns Institutional Information and IT Resources one of four Protection Levels based on the potential impact of unauthorized disclosure or modification. UC San Diego has approved TritonGPT and TritonAI Harness for use with information up to P3 when the use is within an approved service or setup. P4 data is not approved.
Low
Internal information that is generally not public. Unauthorized use or loss could cause minor harm, financial loss, or privacy impact.
Moderate
Information whose compromise could cause moderate harm, privacy impact, financial loss, or legal action. Examples include student education records, UC personnel records, and some personally identifiable information.
High
Information whose compromise could cause significant harm, regulatory action, or civil or criminal penalties. Examples include protected health information, payment-card data, Social Security numbers, and controlled government information.
Read the UCOP classification page Open the UC classification guide (PDF)
Approval up to P3 does not grant access to information or approve every use case. The person or office responsible for the information still determines its classification, permitted use, access, and any additional controls.
The shared foundation
What the foundation separates
Hosting, model access, information sources, and actions solve different problems. Each needs its own approval and operating boundary.
Approved hosting choices
A service can use approved enterprise cloud models or UC-hosted open models running at the San Diego Supercomputer Center. Which route it takes depends on the data involved and the controls that data requires.
- AWS
- Microsoft Azure
- Google Cloud Vertex AI
- UC-hosted models
One gateway for model access
Supported applications reach models through a single managed gateway. It gives everyone the same technical path. It does not approve what data a service may send through it.
Approved sources at request time
When someone asks a question, the assistant looks the answer up in sources it is cleared for. The model itself was not trained on private campus content.
Bounded tools and actions
Skills and connectors expose only the capabilities approved for a service. The service defines who may use them, where a person reviews the result, and when the workflow must stop or escalate.
Where a service runs
Hosting follows reach and risk
A build moves to a more managed home as more people rely on it, as it touches more data, or as failure starts to cost something. The Build overview explains what each rung needs.
- Prototype
- Your own workstation or a sandbox, at
localhost. Fine for learning. Nothing other people depend on. - Team workflow
- A department-owned application on the approved campus application path with campus sign-in, at
apps.ucsd.edu. - Campus service
- A TritonAI or ITS-managed path operated by a named team, at
tritonai.ucsd.edu. - Enterprise service
- The enterprise platform with identity and service management behind it, at
ucsd.edu.
Where people experience it
Where you actually meet it
The same architecture sits behind all of these. Each one still needs its own named owner and an approved support path.
TritonGPT
A campus assistant for chat, documents, and available models.
Website assistants
An assistant sitting on a department site, answering from that site's content.
Mobile experiences
AI features delivered through approved campus applications.
Teaching and learning
Course tools built around what the instructor wants students to learn.
Developer applications and agents
Department tools and supervised workflows built through supported APIs.