Skip to content
Drejt'sia

Making Kosovo's entire legal system searchable and verifiable

256,000+ court decisions and 2,900+ laws, semantically indexed, citation-linked, and queryable with AI. Zero hallucinated sources, by architecture.

Drejt'sia | Making Kosovo's entire legal system searchable and verifiable project screenshot

The challenge

Kosovo's legal system had no structured, searchable foundation. Laws referenced each other inconsistently, court decisions were scattered across portals, and there was no way to trace how a ruling connected to the legislation behind it. Legal professionals had to rely on memory and manual search across dozens of fragmented sources.

Our approach

We built a structured-retrieval database over Kosovo's entire public law corpus, then placed AI on top as the interface layer. Every answer the system gives is grounded in retrieved sources — the LLM cannot invent citations. We fine-tuned our own Albanian legal embeddings, built a citation graph linking laws to decisions and decisions to statutes, and implemented a 6-track hybrid retrieval pipeline to maximize accuracy.

What we delivered

Full-text indexing of 256,000+ court decisions and 2,900+ pieces of legislation
Hybrid search: lexical, semantic, and citation-graph retrieval combined
Architecturally enforced citations — every claim traced to its source document
Citation graph explorer linking laws, amendments, and court decisions
AI Q&A with streaming responses and multi-turn conversation support
Temporal queries for state-of-law as of any historical date
Fine-tuned Albanian legal embeddings with 19-point accuracy improvement over baseline
Multi-language support for Albanian, Serbian, and English legal documents
GDPR-compliant redaction for sensitive court decision data
50-question legal QA benchmark suite gating every production deployment

Tech highlights

Next.js frontend with FastAPI backend and Python 3.12PostgreSQL + pgvector with HNSW indexing for 835K+ embedded chunksCustom fine-tuned legal embeddings (LoRA on multilingual-e5-large)6-track hybrid retrieval: structural, lexical, vector, HyDE, citation graph, rerankingAnthropic Claude + Google Gemini dual-model fallbackD3.js citation graph visualizationOpenTelemetry observability with Prometheus metrics and structured loggingContainerized deployment with Traefik and MinIO document storage
We built Drejt'sia because Kosovo's legal system deserved to be machine-readable. The same rigor we apply to every product: structured data first, AI second, hallucinations never.

Atlas Studio

Drejt'sia

Like what you see?

Let's build something just as good, or better, for you.

Start a project