One front door, many agencies, one platform.
Every module shares the same agencies, agents, services, knowledge, consent records, approval engine and audit trail — so the answer a resident gets, the agents that produced it, what was shared and who approved it never drift apart.
Resident experience
- Front door
The conversational front door residents use on the web, as an installable app, in a widget on any agency’s site, by SMS, WhatsApp or voice, or at a lobby kiosk. It understands requests that span agencies, answers with sources, fills in forms and hands off to a person with the conversation attached.
- Services & journeys
A shared service catalog (Open Referral HSDS compatible) across every agency in the federation, the intents residents express, and life-event journeys — moving, starting a business, buying a home, recovering from a flood — as checklists that cross agency lines.
- Forms & fulfillment
Agency agents gather what a form needs, pre-fill it from what the resident already said, explain every field they ask for, and submit only to a sandbox store once the resident confirms. Production destinations are recorded for later phases and blocked in code.
Agents & federation
- Agent mesh
The registry of agents: the Orchestrating Agent and every agency agent, each with its agent card, skills, hosting and owner. Agents talk over the open Agent2Agent protocol and reach data through read-only Model Context Protocol tool servers; any framework can join by passing the conformance suite.
- Federation & data sharing
Agencies in the federation, the data-sharing agreements between them, and the flows each permits — by data category, purpose and retention. Every time the orchestrator delegates, it passes an agency only the fields an executed agreement allows and the resident has agreed to share.
- Knowledge
Each agency’s allow-listed pages, documents, open data and GIS layers become passages its agent retrieves and cites. Review dates, freshness, a multilingual glossary and content gaps — what residents ask that nothing answers — keep the knowledge ready.
Quality & operations
- Quality & safety
Guardrails on every message in and answer out — emergencies, prompt injection, personal numbers, advice the front door must not give, grounding — plus evaluation sets in every language run on every change, and a review queue where subject-matter experts check flagged answers.
- Live operations
De-identified conversations with their intents, agencies and outcome; every A2A task with its timing; handoffs waiting for staff with their summaries; and alerts when an agent slows down, fails, or accuracy drops.
Planning & insights
- Federation planning
The Phase IB planning package as a working tool: a baseline readiness assessment of every partner’s environment, a sequenced roadmap with dependencies across agencies, a cost model for future phases and a risk register — ready to present to executives and councils.
- Dashboards & reports
Volume, resolution, handoffs, satisfaction and language mix; journeys from first question to submitted form; agent latency and grounding; content gaps — for program staff, agencies and executives, with every chart also as a table.
Administration
- Administration
Organization settings, users and roles with agency scope, approval workflows, segregation of duties, retention and the tamper-evident audit log.
- Integrations
Content and GIS sources, CRM and case systems for handoffs, messaging channels, identity providers, analytics, Sepal Assure and Sepal Learn — through scoped API keys, signed webhooks and an API gateway with rate limits.
Underneath every module
| Area | How Sepal Concierge handles it |
|---|---|
| Vendor-agnostic architecture | Services and agents talk over open standards — A2A between agents, MCP to tools and data, OpenTelemetry for traces. The model provider is swappable per organization (Gemini, Claude, or open-weight models on premises), and partner agents can be built in any framework that passes the conformance suite. |
| Read-only integration layer | Agents reach systems only through tool servers; live and production systems are read-only by declaration and by enforcement in the client. Forms submit to sandbox:// destinations through a single write-back guard. |
| Cloud, on-premises or hybrid | Sepalstack’s multi-tenant service on Google Cloud (us-south1), a dedicated environment, the organization’s own cloud project, or the agent runtime container on the organization’s Kubernetes — mixed per agency. |
| Scale | Stateless functions and containers that autoscale, conversation memory with expiry, cached directory and indexes, parallel delegation with per-agent timeouts. |
| Security | Role-based access checked on the server for every request, agency scope, TLS 1.2+ in transit and encryption at rest, OAuth client credentials or mTLS between agents, scoped API keys, rate limits, dependency and image scanning. |
| Identity | Staff sign in with SAML or OIDC single sign-on (Okta, Entra ID) with MFA. Resident identity is out of scope now; sessions are anonymous and the components accept a verified identity token later (Login.gov or the state’s IdP). |
| Accessibility | Built to WCAG 2.2 AA (WCAG 2.1 AA minimum): keyboard and screen-reader tested, live-region announcements, contrast-checked themes, larger text and high contrast, every chart with a table. |
| AI governance | Grounded answers with citations, low-grounding review, deterministic guardrails for injection and emergencies, evaluation sets in every language, prompt and model versions recorded per task, AI disclosure, and evidence for Sepal Assure. |
| Data protection | Data minimization by design: agencies receive only the categories an executed agreement and the resident’s consent allow; conversations are de-identified before storage; session memory expires; no resident profile is kept. |
| Audit | Append-only audit log of configuration changes, approvals, exports and reveals of personal data, sealed daily with a hash chain. |