Casefinch
AI legal research for the island of Ireland — a 30,000-judgment national corpus, the first computed “is this still good law?” citator for Irish and Northern Irish case law, and paragraph-grounded AI. Built by Amplir end to end, from raw court documents to a revenue-ready SaaS.
The gap
Neither Westlaw nor Lexis ever built a citator for Ireland or Northern Ireland. Practitioners check whether an authority still stands by hand — or not at all — while generic AI confidently fabricates Irish case law. The market is real (≈19,400 practitioners across the island), underserved, and structurally ignored: too small for the incumbents' cost base, too distinct for their English products.
Two legal systems, one product
Distinct courts, citation conventions, publishers and licensing regimes for Ireland and Northern Ireland — unified behind one corpus schema, one citation graph, and a bilingual English/Irish interface.
Trust is the product
Lawyers are trained to distrust unsourced claims. Every AI feature is grounded to the paragraph level and refuses to answer when sources are insufficient — the anti-fabrication discipline is the differentiator, not a disclaimer.
Full-stack delivery
Corpus engineering, information retrieval, graph computation, LLM orchestration, product, billing, and zero-downtime operations — one architect, one coherent system, live and taking payment.
Engineering deep dive
Four decisions that shaped the system — and would shape yours.
Evaluation before architecture
Before writing product code, we built a 900-query gold set (human-reviewed) and measured everything against it. The incumbent approach — a 27-million-vector legal-BERT index — scored below a Postgres keyword search: recall@10 of 0.093 against a 0.109 baseline. It was deleted without ceremony.
The bakeoff selected BGE-M3 dense retrieval (0.513) with cross-encoder reranking over the top-50, embedded at paragraph level with fp16 vectors under an HNSW index. Retrieval quality is a measured property of the system, not a vibe — every model change re-runs the same harness.
The computed citator
The crown jewel. Every judgment’s outbound citations are extracted at the paragraph level, resolved against the corpus, and classified — followed, applied, distinguished, doubted, overruled. The result is a 143,741-edge treatment graph that computes what KeyCite and Shepard’s employ editorial teams to curate, refreshed nightly as new judgments land.
The graph powers the product surface end to end: status badges on every search result, a free citation checker, authority monitoring (“your case was doubted this morning”), statutory-provision stability signals across 74,338 legislation references, and an advocacy-training engine that hides adverse-treatment traps in opposing counsel’s bundle.
National-scale data engineering
The corpus is ingested directly from the official publishers of both jurisdictions — DOCX and PDF, not tidy APIs. Extraction reconstructs numbered legal paragraphs from raw document text, captures judgment dates, and strips artefacts; loading reconciles every record against the existing corpus by neutral citation, so re-ingestion updates in place and never duplicates.
The pipeline is idempotent end to end: a nightly systemd-timed delta sweeps both publishers, embeds new paragraphs, classifies new citation edges, recomputes statute references, and refreshes the public citation reports — unattended. When we re-derived 26,150 judgments in a single weekend migration, users never saw downtime; when 1.87M paragraphs needed re-embedding, a rented RTX 4090 did it in ninety minutes for about two dollars.
Provenance as architecture
The first question sophisticated buyers ask a legal-data company is “where is your corpus from, and are you licensed?” Casefinch answers with a database column. Every case row carries its source and licence; the corpus was migrated from scraped aggregator data onto the official publishers’ own reuse terms — the Courts Service of Ireland’s commercial reuse licence and Crown copyright under the Open Government Licence — with 85.6% of rows clean-licensed in one weekend and the remainder confined to pre-2001 history.
Compliance engineered as data, not asserted in a policy document: attribution is rendered from the same columns, and the ongoing pipeline only ever touches clean sources.
One graph, six products
The corpus, the vectors, and the treatment graph compound into a product family — each feature makes the others harder to replicate.
Research
Semantic search over 1.99M paragraphs with reranking and treatment badges inline.
Citation checker
Paste any citation; see how later courts treated it. Free — the acquisition wedge.
Grounded Q&A
Composed answers pinned to real paragraphs; refuses when sources are insufficient.
Legislation
“Which cases interpret this section?” — 74,338 statute references across 4,240 acts, with provision-stability signals.
Authority monitoring
Pin the cases a live matter relies on; get alerted when the ground shifts.
Weekly digest & advocacy
An AI-drafted, editorially-gated weekly digest of new law, and structured advocacy practice against real decided cases.
Stack
This is what we build when the client is ourselves.
Casefinch is Amplir’s own product — which means every decision above was made the way we’d make it for you: evaluation before architecture, compliance in the schema, pipelines that run unattended, and AI that shows its sources. If you have a data-heavy or AI product that needs to actually work, this is the standard it arrives at.