One platform from first test to release decision.
Every module shares the same agents, versions, knowledge sources, tests, approval engine and audit trail, so a score, the evidence behind it and the decision made on it never drift apart.
Build & test
- Agents & environments
Register chatbots, virtual assistants, task agents and multi-agent workflows — built in-house or supplied by a vendor — with their environments, versions, risk tier and how Sepal Assure reaches them: API, Bot Framework Direct Line, Microsoft Teams, a browser, transcripts or traces.
- Knowledge sources
Connect SharePoint sites, document uploads, websites and repositories. Documents keep their headings, version, effective date and review date, and are split into passages that tests, citations and grounding checks point at. Sepal Assure also finds the documents that disagree with each other.
- Test design
Suites of single-turn prompts, scripted and simulated multi-turn conversations, task tests with expected tool calls and outcomes, adversarial attacks, fairness pairs and language-parity tests — generated from approved knowledge or built from templates, and approved by subject-matter experts before they run.
Evaluate
- Evaluation
Rule checks first, meaning-based comparison second, calibrated judges third. Every answer is scored on accuracy, completeness, relevance, clarity, consistency, grounding, citations, policy, safety, task completion and trajectory — each with its reason — and judges count only where they agree with your experts.
- Runs & regression
Run suites on demand, on a schedule or from a pipeline, against many agents and projects at once. Every run records the version, model and environment; repeated trials give confidence intervals, so a regression is a real change and not model noise. Load tests replay whole conversations and score quality under load.
Operate
- Production monitoring
Score a masked sample of real conversations, run synthetic probes against production, and fingerprint each agent’s behavior every day so a provider’s silent model update is caught and measured. Alerts go to email, Teams and ServiceNow.
- Findings & releases
Failures become findings with their evidence, linked to ServiceNow, OpenText ALM or Azure DevOps. Release reviews assemble the gate result, the comparison with the baseline, open findings and accepted risks, and route them to your approvers. AI incidents become permanent regression tests.
Govern
- AI governance
An inventory of the agency’s AI systems, the controls they answer to (NIST AI RMF, the Texas Responsible AI Governance Act, OWASP Top 10 for LLM applications, ISO/IEC 42001), and evidence collected automatically from runs, calibrations, reviews and monitors — ready as a packet for auditors, the board or the state.
Insights
- Dashboards & reports
Portfolio, agent, trend, comparison, review and performance dashboards built for technical and non-technical readers, with drill-down from every number to the response behind it. Reports export to PDF, Excel and CSV, arrive by email on a schedule, and feed Power BI.
Administration
- Administration
Organization settings and evaluation defaults, projects, users and roles with project scope, approval workflows, segregation of duties, retention and legal holds, budgets and the tamper-evident audit log.
- Integrations
ServiceNow, OpenText ALM and UFT, Azure DevOps, GitHub Actions, GitLab and Jenkins quality gates, SharePoint through Microsoft Graph, Teams alerts, Power BI, SIEM export, Okta SSO and SCIM, runners, API keys and signed webhooks.
Underneath every module
| Area | How Sepal Assure handles it |
|---|---|
| Identity federation and SSO | Okta as the identity provider over SAML 2.0 or OpenID Connect; scoped, expiring API keys for pipelines and in-network runners; SCIM 2.0 provisioning; MFA enforced by Okta, with an optional second factor required again at sign-in; no vendor-specific identity platform. |
| Role-based access | Roles at organization, project and agent level, checked on the server for every request; custom roles from individual permissions; segregation-of-duties rules. |
| Audit logging | Append-only audit log of sign-ins, permission and configuration changes, test edits, approvals, runs, review decisions, exports and value reveals, sealed daily with a hash chain and streamable to a SIEM. |
| Data residency and training | All data stored and processed in US regions (Dallas primary); model calls only through US endpoints; customer data never used to train models. |
| TX-RAMP | Controls aligned to the NIST SP 800-53 Moderate baseline used by TX-RAMP Level 2; underlying Google Cloud services hold TX-RAMP authorization. |
| Accessibility | Built to WCAG 2.2 AA: keyboard and screen-reader tested, contrast-checked, every chart has a table. |
| Availability and recovery | 99.9% monthly target; point-in-time database recovery; daily backups to a second US region; RPO 1 hour, RTO 8 hours, tested yearly. |