What this is
A unified, multilingual AI portal we designed, built, and own outright as part of our internal capability portfolio. One platform, two brand skins; it serves a county-government audience and a K-12 school-district audience side by side, with shared infrastructure underneath and tenant-aware separation throughout.
The build was scoped against a public civic-government RFP for an AI portal serving both a county and its public schools. We share architecture and the working system in qualifying conversations under NDA. A live demo is available on request.
Why we built it
Local-government technology has a hard set of constraints that most off-the-shelf AI products do not respect together: residents speak many languages; content lives across many sites and document types; accessibility is not optional (WCAG 2.1 AA and PDF/UA are the floor, not the ceiling); public-records law requires every model interaction be auditable; and US data localization, no-training-on-customer-data, and US-developed-AI requirements are increasingly written into civic solicitations.
We wanted a single portal that respected all of those constraints from the first commit, not as an after-the-fact certification exercise.
What it does
- Multilingual semantic search across public agency content, with retrieval grounded in a structure-aware document index (sections, headings, page anchors) so citations land on the right paragraph, not just the right page.
- Topic-specific chatbots for civic use cases (building permits, citizen request lines, county codes, school-board policies) sharing one RAG core but presenting under their own narrow scope and system prompt.
- WCAG 2.1 AA + PDF/UA-1 accessibility audit and remediation for the same documents the chatbots cite. Operators can drop in a PDF and get back a tagged, screen-reader-ready PDF plus an audit report. Embed mode drops an “Audit this PDF” affordance onto any third-party page next to existing PDF links.
- Natural-language insights over public datasets, with the conversational layer cross-linking back to citations in the search and chat surfaces.
- Embed widget loader so any agency page can drop in a small script tag and get search, chat, or audit UI without rebuilding their CMS.
- Audit, segregation, and localization built in: every model call writes a structured audit row; tenant identifiers flow through every business write; AI providers are constrained to a US-developed and US-hosted whitelist; non-US-origin vision models are isolated as benchmark-only and excluded from production defaults.
Built with
| Layer | Stack |
|---|---|
| Backend | Python, FastAPI (async ORM with migrations) |
| Frontend | Vite + React + TypeScript + Tailwind CSS, with embed-widget loader |
| Database | PostgreSQL with vector + full-text extensions |
| Object storage | S3-compatible |
| LLM | Frontier LLM from a US-developed, US-hosted provider |
| Embeddings | Open-source embedding model |
| Document AI | Pluggable layout-aware document-understanding backends with origin-aware compliance flags |
| Internationalization | Server-side multilingual prompts plus UI translation framework |
| Accessibility | In-repo PDF tagging, structure-tree generation, and WCAG/PDF-UA validation |
| Reproducibility | Docker Compose (clean clone to running stack) |
What this means for you
If you operate a public-facing service where residents, students, families, or constituents need to find, understand, and act on information across many documents, many topics, and many languages, and you have to do it under public-records, accessibility, and US-data-localization discipline, we can:
- Build a unified portal that serves multiple audiences under their own brand skins on one tenant-segregated stack.
- Combine semantic search and topic-specific chatbots so users get conversational access without losing source citations.
- Deliver WCAG 2.1 AA and PDF/UA-1 audit and remediation as a first-class feature, not a bolt-on.
- Engineer audit logging, tenant segregation, and US-developed-AI enforcement in from day one, with compliance design that maps cleanly to civic-procurement RFP requirements.
- Provide an embeddable widget mode so agency teams can light up AI surfaces in their existing CMS pages without a platform replacement.
Want to see a live demo or walk through the architecture under NDA? Contact us.