Source-grounded legal AI. Zero hallucinated citations.

The challenge

Norwegian family law is dense, multilingual, and high-stakes. Families facing Barnevernet (child welfare) proceedings or custody disputes need accurate legal information — but the stakes of getting it wrong are severe. Standard LLMs hallucinate case citations and misstate procedural rules, which is not merely unhelpful in a legal context: it can cause families to miss filing deadlines, pursue invalid appeals, or misunderstand their rights at critical moments.

Do Better Norge's users span four languages and four distinct legal frameworks: Norwegian statutes (Barneloven, Barnevernloven), ECHR Article 8 precedents from Strasbourg, Hague Convention cross-border abduction cases, and Barnevernet administrative appeal procedures before Statsforvalteren. Existing legal search tools offered no coverage of Norwegian family law at this depth. There was no product on the market that could give a Ukrainian-speaking parent, in plain language with proper citations, an accurate explanation of how to appeal a Barnevernet care order.

The platform also serves journalists and legal professionals who need to verify claims against primary sources — making citation accuracy a non-negotiable requirement, not a nice-to-have.

The solution

Gilligan Tech Inc. designed and deployed a multi-layer AI intelligence stack on the Gilligan Tech Inc. AI Platform, delivered as a fully managed service powering both dobetternorge.no and dobetterlegal.com. The stack is built around one principle: every answer must be traceable to a specific source document. The AI does not speculate; it retrieves, explains, and verifies.

The core is a 220,000-chunk legal corpus embedded into Qdrant using nomic-embed-text, feeding a 4-stage pipeline where a Planner decomposes the user's question, a Researcher retrieves ranked source chunks, an Explainer generates a source-grounded answer with Kilder: (Sources) citations, and a Verifier checks the answer for overclaiming or unsupported authority before it reaches the user.

On top of this, Gilligan Tech Inc. built and hosts a CaveauAI MCP server exposing 15 specialized legal tools — from Barnevernet case analysis and advocate brief generation to citation graph construction and multilingual translation — that advocates and legal professionals can use directly from their AI tools.

Implementation

1. Legal corpus build: 220,000+ chunks ingested from Norwegian statutes (Barneloven, Barnevernloven), ECHR decisions on Article 8 (right to family life), Barnevernet procedural guidance, court decisions, ombudsman reports, and Hague Convention documentation. All documents embedded with nomic-embed-text (768-dim) into a dedicated Qdrant collection.
2. Custom LLM fine-tuning (KaaP): Gilligan Tech Inc.'s Knowledge-as-a-Product fine-tuning pipeline distilled Norwegian family law expertise from a NorwAI-24B teacher model into a compact Qwen 3 8B base. The resulting model — dobetter-norge-v4 — answers Norwegian law questions with domain accuracy a general-purpose model cannot approach, while running at inference speeds suitable for real-time use.
3. 4-stage verification pipeline: Every user query runs through four stages: a Planner decomposes the question into targeted legal research queries; a Researcher retrieves ranked, source-cited chunks from the corpus; an Explainer generates a structured answer with full Kilder: citations; and a Verifier checks whether the answer overclaims, cites unsupported authority, or exceeds what the retrieved sources actually establish. Flagged outputs are revised before delivery.
4. MCP server (15+ tools): Gilligan Tech Inc. built and hosts a CaveauAI MCP server exposing specialized tools: corpus search, legal analysis, Barnevernet case analysis, advocate brief generation, case workbench planning, citation graph, timeline construction, document summarization, redaction, and multilingual translation. The server is available to Pro-tier platform members and legal professional subscribers via the MCP protocol.
5. Multilingual delivery: All AI features operate across Norwegian, English, Ukrainian, and Polish via LiteLLM routing and language-aware prompting. This directly serves immigrant families — particularly Ukrainian and Polish communities — who face the highest barriers to accessing Norwegian legal information in a language they can act on.

Results

Technology