Model and workload fit
Use the smallest capable model, reduce repeated work, and evaluate whether retrieval, rules, or a simpler tool can solve part of the task.
Service design and operations
Use the resources the work requires
Sustainable AI starts with choosing an appropriate model and workflow, avoiding unnecessary processing, reusing shared capabilities, and measuring quality, latency, cost, and resource use together.
No hosting choice is automatically best for every workload. TritonAI revisits the route as models, demand, evidence, and campus requirements change.
Design decisions
What teams should account for
Shared infrastructure
Reuse the gateway, knowledge sources, skills, connectors, and hosting patterns campus already operates.
Service lifecycle
Monitor use and outcomes, revisit inefficient routes, and retire stale indexes, duplicated context, services, or pipelines that no longer justify their operating cost.
Hosting context
UC-controlled and enterprise cloud routes
TritonAI supports approved enterprise cloud models and open models on UC-controlled infrastructure, including resources at the San Diego Supercomputer Center. The service, data, required capability, and approved controls determine which route is appropriate.
UC-controlled hosting can support shared capacity and local operational choices. Enterprise cloud routes can provide capabilities and scale under institutional agreements. Environmental comparisons still require workload-level evidence.
Review the whole workload
WorkloadTask, data, demand, and controls
- Quality
- Can the route complete the task reliably?
- Latency
- Does it respond within the service need?
- Cost
- What does the full service require to operate?
- Resources
- What compute, storage, retrieval, and repeated processing does it use?
Operating practice
Questions for every service review
- Is the model larger than the task requires?Compare a smaller approved model and a non-generative method where either could meet the need.
- Are we repeating work?Reuse approved sources, cache safe results where appropriate, and remove unnecessary retries or duplicate pipelines.
- Is shared context still current?Check the source, owner, review date, index, and cached material. Correct or retire stale context before it causes more work.
- Does the evidence justify continued operation?Review use, outcome, reliability, support effort, and resource consumption together.
- Can another team reuse this?Package stable instructions, integrations, and evaluation patterns so the next service starts further ahead.