PRISM: SaaS · AI platform case study
An AI observability platform that catches problems before your users do.
Overview
PRISM is an AI observability and evaluation platform that catches problems before your users do. I designed the full product experience, from the first login screen to the daily workspace.
The problem
When AI agents start handling real conversations, such as support, sales and onboarding, one bad response can break trust instantly. Teams need to see what their agents do, and stop risky or malicious behaviour before it reaches the end user.
Goals
Structure of the product
One left navigation groups the work by what a team does: Observe (traces, agent intelligence, sessions, end users, agent runs, remediation), Evaluate, Project and Integrate (connectors, model inventory, data export). A workspace switcher and credit balance stay in the top bar.
Sign in
A clean auth screen with Google and GitHub sign-in, and social proof from a real founder quote to build trust from the first screen. A dark, textured panel carries the product promise and three headline facts.
Dashboard
Six KPIs sit at the top: total traces, active agents, average latency, guardrail blocks, compliance score and cost today, each with a small trend. Below, a compliance score trend and the agent fleet table show how each agent is doing, so the day starts with one glance.
Traces
Every AI interaction is logged and searchable. Filters for date range, agent, status and latency narrow the list, and blocked or flagged traces are marked in the row. Teams can import historical traces or export to CSV.
Agent runs
A list of runs beside a graph of the selected one, following the path Input, Retrieval, LLM, Output with status, duration and tokens per step. Tabs switch between graph, timeline, architecture and evaluation, and a run can be downloaded as JSON, copied as schema or re-evaluated.
Design approach
A dark hero paired with a clean light workspace, so the product feels serious but never overwhelming. One green accent marks status and primary actions; everything else stays neutral.
Conclusion
PRISM turns agent monitoring into a handful of calm screens that follow how a team works: sign in, check the dashboard, drill into traces, inspect a run.