Skip to main content

About TritonAI

TritonAI Strategy

TritonAI makes AI available across UC San Diego as a shared campus capability, then helps teams turn specific, recurring work into supervised and measurable services. The same foundation supports chat, embedded assistants, developer applications, and agentic workflows.

Campus impact at a glance

  • 16Campus websitesPowered by the TritonGPT widget
  • 91%Faster contract review120 to 11 minutes in the reported workflow
  • 99.86%Transcript match accuracyPublished validation sample of 3,700+ records
  • 81%Helped understandingInstructional AI pilot survey response

Each result covers only the service and measurement period shown. None of them is a campus-wide average.

Six principles

Trust is infrastructure

Choose approved UC-hosted or enterprise cloud routes according to the service, data classification, capability, and required controls. Governance covers the model and every other part of the workflow.

Make AI a utility

One shared setup, one way in, and one place the costs show up. Teams start faster and can copy what already worked somewhere else.

Solve specific pain points

We go after the jobs people complain about: contract review, search, scheduling, accessibility, instructional support. A tool built for one of those pays off faster than general-purpose chat.

Meet users in the workflow

AI shows up inside Blink, public websites, teaching tools, the mobile app, and departmental systems. People use it when it is already where they are working.

Stay model-agnostic

Services are built on the shared gateway and the evaluation and governance around it, so swapping the model underneath does not mean rebuilding the service.

Prepare for agency

Assistants answer questions. Agents go further: they fetch data, call tools, and finish multi-step work under supervision. The skills, APIs, connectors, and monitoring that make that safe are what we are building now.

How an idea becomes a service

Discover

Name the people, workflow, approved data, service owner, review point, and result that should improve.

Pilot

Use a bounded environment, realistic examples, and close human review to test whether the approach fits.

Prove

Measure reach, efficiency, quality, and trust. Refine the workflow until the evidence supports wider use.

Operate

Establish approved hosting, accessibility, monitoring, and support. Name the escalation path and service owner before people depend on it.

Where AI meets the workflow

TritonGPT

The full chat experience. Sign in with your campus account, pick a model, upload documents, and use assistants that already know about campus.

Open TritonGPT

UC San Diego app

The campus Assistant answers questions inside the mobile app, using approved context available to that experience.

Learn about the app

Website widget

Approved sites embed an assistant grounded in their public content. The public evidence package tracks where it is live.

Explore the widget

Developer APIs

Departments add approved model access to their applications through the shared gateway and its accountability controls.

Explore the APIs

TritonAI Harness

The primary supported workspace for building agentic applications with campus model access, skills, connectors, and review gates.

Explore the Harness

Skills Library

Reusable instructions and patterns let teams adapt work that has already been tested without rebuilding every integration.

Browse the Skills Library

How the work changes

Answer

A person asks a question, provides context, and checks the response. TritonGPT and embedded assistants support this interaction.

Act

A supervised workflow works toward a goal, using approved skills, sources, and connectors across several steps with review gates.

Improve

The service owner reviews the trace and outcome, corrects the context, and promotes tested patterns that another team can reuse.

Models will keep changing. The service depends on approved integrations and clean sources. Reusable skills, honest evaluation, and named ownership keep it running.

Shared context follows the same operating discipline. A person approves what becomes reusable, and the record carries its source, owner, review date, and freshness expectations. A service retrieves only the context its task permits and presents proposed changes for review. The trust architecture explains the shared service foundation.

Explore use cases View the roadmap