Public-sector AI has a different bar. A shopping assistant that gets a detail wrong costs a sale. An assistant that misstates the law, quotes a Bill as if it were already enacted, or leaks someone's ID number costs trust. When we built our AI knowledge platform, we designed for that bar from the start.
One core, scoped per request
Most organisations that want an AI assistant end up with a separate bot, a separate vendor contract and a separate pile of documents for every department. We took the opposite approach.
There is one AI core and one knowledge store, partitioned by namespace. What a request can see is decided by its API key, never by what the caller claims. A public assistant sees the public corpus. An internal staff assistant sees its department's private documents plus the public ones. Onboarding a new institution does not change the core: it gets its own keys, its documents go into its own namespace, and its assistant points at the same API.
Only certified knowledge gets in
The platform separates authoring from serving. Laws, policies, plans and service catalogues are ingested, structured and certified by a human before anything is published. Only certified final documents reach the AI's knowledge store. The AI never sees drafts or work in progress.
Ingestion itself follows a simple rule: check with deterministic rules before asking a model. Blank pages are detected and skipped. Every page transcription is validated before it is saved. The document's own structure (articles, sections, clauses) is parsed and becomes the backbone everything else hangs on. For Bills, the operative text is separated from the explanatory memorandum, so the AI never confuses commentary with law.
Search that reads like a lawyer
When someone asks a question, our retrieval engine does far more than a simple similarity search:
- It expands acronyms and defined terms using the documents' own glossary.
- It combines meaning-based search with exact keyword search, so section references and defined terms are never missed.
- It reranks results for precision and gives binding law priority over circulars and guidelines.
- It brings in amendments, repeals and superseding documents automatically when you land on the original Act.
- It can answer "what did the law say on this date" thanks to versioned, dated content.
- It widens a matching clause to its full section, so the AI reads the whole provision, provisos included.
Each of these steps can be tuned or reordered per assistant without a code change, and every step leaves a trace, so every answer can be audited.
Answers that get cheaper over time
Not every question needs a large language model. Our engine tries the cheapest correct path first: a cached answer, then a library of human-approved answers, then a fully grounded AI answer with citations, and only then a human agent. When an agent resolves a question, that answer can be reviewed and added to the approved library. Quality goes up and cost per question goes down as usage grows.
Safety built in, not bolted on
Every assistant inherits the same guardrails, and a client can only make them stricter:
- Answers are grounded in certified sources with citations, or the assistant declines.
- Personal information is protected on the way in and redacted on the way out.
- Moderation and prompt-injection defences run on every message.
- Low-confidence answers are handed to a human.
- Guardrails run independently of the AI model, and fail safe.
No vendor lock-in
All AI traffic passes through a single gateway. Changing or mixing AI providers, including a self-hosted model for sovereign deployments, is a configuration change. Each task is routed to the most cost-effective capable model, with fallback and per-client budgets. Your knowledge and your workflows stay yours.
Built for institutions with a lot to say
If your organisation has policies, regulations, service catalogues or legal instruments that staff and the public struggle to navigate, this platform was built for you. Ministries, county governments, regulators and large enterprises can run public assistants, internal staff assistants or a search API on the same core.