Program overview
TritonAI
A shared AI platform for service, instruction, development, and everyday campus work.
TritonAI includes TritonGPT, campus data tools, a model hub, developer APIs, embedded assistants, reusable workflows, and early support for supervised agents. These services share a common governance and infrastructure model.
UPDATED
September 2026
CONTEXT
University of California, San Diego
Executive summary
Scope of the program.
TritonAI is UC San Diego's institutional AI program, serving 73,000+ students, faculty, and staff. It includes infrastructure, task-specific assistants, campus data tools, instructional pilots, developer access, and shared workflows.
TritonGPT is the shared platform at the center of that work. The SDSC-hosted environment is available to students, staff, and faculty and supports model choice, on-premises and approved cloud routes, developer APIs, embedded website support, and agent workflow components. TritonGPT's default model now runs on UC San Diego-hosted infrastructure, keeping routine campus AI traffic on UC-controlled hardware.
The platform has also been licensed to peer institutions, including UC Berkeley (BearGPT), UC Agriculture and Natural Resources, Fairleigh Dickinson University (FredGPT), and UC Merced (SCOUT), which opened to its full campus in September 2026.
01
Knowledge
Curated campus context, uploaded documents, and live data tools
02
Model Access
Self-hosted and approved cloud models through governed routes
03
Developer Surface
APIs, credits, keys, endpoints, and usage reporting
04
Workflow
Assistants, embedded support, tools, and reusable agent scaffolding
Program parts
What TritonAI Includes
TritonAI includes the shared pieces needed to support campus adoption: a governed platform, embedded experiences, instructional tools, live data integrations, developer access, reusable skills, and a path toward supervised agent workflows.
TritonGPT Platform
A secure, SDSC-hosted foundation for campus AI with assistants built around specific university workflows and campus-aware context.
Learn more
Embedded Website Support
Embeddable AI support for campus websites, now paired with live campus data patterns such as parking availability and location busyness.
Learn more
Instructional AI
Course-grounded student and instructor assistants built around instructor-approved materials and teaching workflows, now natively integrated with Canvas.
Learn more
Shared Resources and Adoption
Guides, prompting practices, examples, and training that help campus teams use TritonAI more effectively.
Learn more
Developer APIs
A governed gateway to self-hosted and approved cloud models, with supported build paths including the TritonAI Harness, n8n workflow automation, and developer APIs across four hosting lanes that carry a project from a personal sandbox to an enterprise service.
Learn more
Tools, Skills, and MCP
A growing toolkit of reusable tools, MCP-based connectors, shared skills, and workflow patterns for common campus tasks.
Learn more
TritonAI Harness
A campus development environment for tool-using AI workflows, now a supported build path for development teams. Early usage includes 22.5B tokens, 119 merged pull requests, and 1,139 AI reviews.
Learn more
Developer API Program
A governed path from a personal experiment to a campus service.
A recurring challenge in campus AI is moving from a working prototype to something the university can operate. The Developer API Program provides shared infrastructure and defined ownership for that process. Before choosing a path, a project defines its user, its task, the approved data it may touch, and the outcome to measure it against.
01
455.3B02
143.9M03
92.6%Campus need
A question, a workflow, a document task
TritonAI Gateway
identity · routing · policy
Approved models
Chosen for the task — UC-hosted or approved cloud
Approved knowledge
Grounded at request time, not trained into weights
Approved tools
Bounded capabilities a person can check
Supported ways to build
All paths route through the TritonAI LLM Gateway, which reaches approved commercial providers and UC-hosted open-weight models across chat, reasoning, vision, image generation, OCR, and coding.
Preferred for most staff
TritonAI Harness
A campus-oriented desktop experience for Mac and Windows, open to anyone with a Gateway key. Bundled installation, UC San Diego model access, campus skills, and Microsoft 365 connections.
Recurring processes
n8n Workflow Automation
Connect applications and APIs around a defined trigger — schedules, webhooks, messages, or file events — for repeatable, reviewable automation.
Technical builders
Claude Code
A command-line environment with direct Anthropic model access, for people already comfortable working in a terminal.
Experienced builders
Codex desktop app
An OpenAI desktop environment for managing code, repositories, parallel tasks, and agent workflows.
Four hosting lanes, chosen by scope and risk
A project does not have to guess how much process it needs. The lane follows the audience and the blast radius, and it can change as the work grows. Whatever the lane, the owner stays responsible for approved data, testing, accessibility, support, and human review.
- Lane 0Personal workspaceIndividual exploration in a desktop or sandbox environment.
- Lane 1Department applicationA focused application with a known audience and a business owner, published through approved campus paths.
- Lane 2Managed campus serviceA shared workflow with integrations or broader operational impact, hosted by TritonAI or ITS.
- Lane 3Enterprise serviceCampus-wide delivery with formal architecture and governance.
Recorded work
TritonAI in Action
Examples from the TritonAI portfolio, built on TritonGPT and connected to everyday university work.
Job Description Helper
Helps HR and hiring managers draft clear, compliant job descriptions in minutes.
Fund Manager Coach
Guides grant and fund managers on compliance, allowable expenses, and spending restrictions.
UC San Diego Assistant
Answers campus policy and service questions using curated departmental knowledge and institutional resources.
Contract Reviewer (In Production)
Automated NDA and contract review via monitored inbox + portal. Auto-redlines third-party contracts against UC San Diego and UC Terms & Conditions and Legal Position.
Measured signals
Early Signals from TritonAI
A few measures that show how the work is affecting administrative workflows, public support, and learning.
01
91%02
83%03
30+04
$500KOperating principles
How TritonAI Is Taking Shape
TritonAI is being developed around a few practical principles: governed infrastructure for sensitive work, shared access across projects, tools designed around specific campus pain points, model choice without fragmented governance, and an architecture that can evolve from assistants toward supervised agency.
Trust Is Infrastructure
Sensitive work in HR, finance, legal, student support, and institutional operations needs governed infrastructure. TritonAI separates local/self-hosted capacity from approved cloud routes so projects can match model choice to data sensitivity, cost, and task complexity.
Make AI a Shared Utility
Shared infrastructure and zero-variable-cost access can make experimentation available to more people. The program aims to provide broad and equitable access to useful tools.
Solve Specific Pain Points
The strongest use cases have been concrete ones: contract review, public support, search, scheduling, performance review alignment, transcript workflows, and instructional support. Focused tools create trust faster than generic chat.
Meet People in the Workflow
TritonAI is showing up in Blink, public websites, teaching tools, mobile experiences, developer projects, and departmental applications. Adoption improves when AI appears where work is already happening.
Stay Model-Agnostic
The architecture is built around gateways, canonical model routes, connectors, usage reporting, and governance rather than a single provider. That keeps campus teams flexible as models, prices, and commercial terms change.
Prepare for Agency
The next step is moving from answering questions to completing supervised multi-step work. APIs, reusable skills, connectors, model observability, and access monitoring are the pieces that make that shift practical.
Assistant index
Examples Across the TritonAI Portfolio
These are not generic chatbots. They are focused tools built around specific service, instructional, research, and administrative workflows where grounded context and institutional trust matter.
01
UC San Diego Assistant
Cross-functional knowledge bridge drawing from 70+ curated department websites, hundreds of ServiceNow knowledge base articles, and UC San Diego policies.
02
General AI Assistant
Writing and editing support without UC San Diego-specific context or reference links.
03
Job Description Helper
Helps HR and managers draft clear, compliant job descriptions much faster.
83% reduction in drafting time
30 min → 5 min average
04
Contract Reviewer (In Production)
Automates NDA and contract review for the UC San Diego Procurement team via a monitored inbox and portal. Cleans formatting, extracts clauses, applies UC San Diego rulesets, and produces annotated redlines aligned with UC Legal Position and university template terms.
91% time savings for NDAs/T&Cs
120 min → 11 min average
05
Fund Manager Coach
Expert guidance on grant compliance, spending restrictions, and financial management for research funds.
06
Phishing Analyzer
Scores suspicious emails for phishing risk and explains warning signs to protect campus data.
07
Deep Research Assistant
Synthesis and analysis across hundreds of research documents and institutional data points.
08
Triton Developer Assistant
On-demand coding support and documentation analysis for staff and administrators.
09
Socratic Tutor
Guided learning assistant that helps students master materials without providing direct answers.
10
Instructor AI Assistant
Automated generation of quizzes, summaries, and study guides directly from lecture content.
11
Internet Search Assistant
Conversational AI with internet access for real-time information and answers.
12
Expert Notetaker
Converts transcripts or notes into structured documentation focused on decisions and actionability.
13
Project Charter Assistant
Helps draft professional project charters.
14
AI Workplace Opportunity Assistant (v0.1)
Interviews users about daily tasks to suggest tailored AI use cases and prompts.
15
Performance Review Evaluator
Helps supervisors and HR teams assess whether performance ratings are supported by written comments, with stricter review of rating/comment alignment.
16
Data Dictionary Assistant
Provides definitions from specific UC San Diego data sources like the Activity Hub and Blink.
17
Supplier Search Tool
Assists in finding campus suppliers and explains the UC San Diego purchasing process.
18
UC San Diego Benefit Assistant
Answers questions regarding employee benefit packages.
19
Student Scheduling Assistant (Live)
AI-powered scheduling assistant launched July 2026, now live on students.ucsd.edu with the Class Planner go-live July 20. Serves 20K-40K students in waves, letting them build optimized schedules through natural-language conversations based on subject codes, availability, and prerequisites.
20
Enterprise Data Agent
Natural-language interface to UC San Diego's enterprise data warehouse. Turns plain-English questions into structured queries across Activity Hub, UCPath, and campus data sources.
21
Transcript Matching (In Production)
OCR and transcript verification workflows for enrollment-related review. Matches student transcripts against program requirements automatically, reducing manual review time for admissions and records teams.
99.86% accuracy in validation
Across 3,700+ records with five errors
22
Dissertation Formatter (Pilot)
Helps students identify which formatting rules their dissertation breaks, with clear guidance on which fixes still need a person to decide.
23
BioBib (Pilot)
Assembles a faculty BioBib draft from approved activity data — publications, teaching, service, and awards — ready for review.
External recognition
Recognition & Awards
Roadmap
Where TritonAI Is Going Next
The next phase expands beyond chat into a broader tools hub: richer model access for developers, more embedded campus experiences, workflow orchestration, observability, live data integrations, and new support for academic and administrative use cases.
01
Instructional AI Expansion
Canvas integration is live, giving enrolled students a course-connected AI Tutor grounded in instructor-approved materials. Next: more course pilots and simpler assistant creation for instructors.
02
Developer API Program (Live)
Live with supported build paths — the TritonAI Harness, n8n workflow automation, and developer APIs — across four hosting lanes from personal sandbox to enterprise service. 455.3B tokens and 143.9M API requests through the gateway in the first eight months of 2026 (january–august), 92.6% on self-hosted models. Next: broader recharge paths from prototype to funded campus application.
03
Mobile Integration (Live)
The UC San Diego Assistant is live in the official UC San Diego mobile app on iOS and Android, delivering campus knowledge, parking availability, and location busyness from the phone.
04
Model and Service Observability
Catalog monitoring, access checks, usage signals, and AI-assisted detection for service anomalies and operational risk.
05
Academic Advising Agent
Proof-of-concept support for tier-one academic advising questions and triage.
06
Agent Builder and Workflow Tools (Live)
The TritonAI Harness is out of pilot and a supported build path for development teams. Low-code orchestration, reusable skills, MCP-style connectors, and guardrails for supervised multi-step campus workflows. 22.5B tokens used, 119 merged PRs, 1,139 AI reviews in first weeks.
Program lesson
What We're Learning Through TritonAI
One lesson so far is that shared infrastructure makes institutional AI easier to support. Secure hosting, model choice, reliable knowledge sources, live tools, reusable connectors, governance, and partnership with the people doing the work all affect the result.
TritonAI is one way of learning how shared institutional AI can become practical and trusted.
Want to Compare Notes?
If you're building something similar in higher education, I'd be glad to compare notes and hear what your institution is learning.
Start a Conversation