The problem: the case file as a labyrinth

A complex legal case is rarely a tidy set of documents. It is a heterogeneous file: contracts in Word, scanned company records and expert reports (often with handwritten annotations), email chains in .msg or .eml format, digitally signed attachments in .p7m, photographs of documents. Dozens or hundreds of files, to be read, cross-referenced and cited correctly. For a professional this means days of low-value work before the real analysis even begins. The same applies to administrative cases — tenders, authorisations, disputes with the Public Administration — where the document dossier grows over years.

Docu Lekton is the software we designed and developed entirely at Lekton.AI to solve exactly this problem: turning the case file into structured, queryable and — above all — verifiable knowledge. It is neither a prototype nor an integration of other people’s tools: it is a complete application, written by our team and already operational on real cases.

From ingestion to knowledge

The system accepts the case file as it is. Each file type follows the most suitable processing path: native PDFs are read directly, scans go through a multi-engine OCR cascade — a component we developed ourselves — that automatically selects the best technology among state-of-the-art vision models (Google Gemini, Google Cloud Vision, Anthropic Claude) with a guaranteed local fallback: this way even difficult handwriting, stamps and degraded documents are recovered. Emails are opened, decoded (including .p7m cryptographic envelopes) and turned into text with their metadata.

The result flows into a semantic index that powers two key functions: the phased guided analysis and the chat over the case, with which the professional can query the entire file in natural language.

The critical point: zero citable hallucinations

Anyone working with legal documents knows that an AI that “invents” even a single detail is unusable. That is why we engineered two distinctive mechanisms into Docu Lekton.

The first is the document manifest: classic semantic retrieval, on files of 50+ documents, risks “seeing” only a fraction of them. Our system builds a summary of each individual file and includes it in every analysis, ensuring that no document in the case file is ignored.

The second is citation verification: every reference produced by the AI is checked against the original text with a multi-strategy matching of our own implementation. If a statement finds no support in the sources, it is not cited. The final report — generated in Word or in share-ready HTML — links every conclusion to the document it comes from.

Under the hood (for developers)

Docu Lekton is a desktop application written in Python, with FAISS vector indexing, a page-level parallel OCR pipeline and automatic generation of reports in Word and HTML. We believe in technical transparency: two of the system’s founding components — the multi-engine OCR cascade and the citation verification engine — are published as open source projects on our GitHub, where anyone can examine their code: ocr-cascade and verified-citations. The domain methodological know-how remains, of course, the proprietary heart of the product.

The results for the professional firm

The practical effect is a drastic compression of preliminary-review times: organising the file, extracting the content and the first analysis go from days to minutes, and the professional gets into the substance with an already-structured picture and traced sources. AI does not replace professional judgement: it puts it in a position to start from the right point.

---

Want to see Docu Lekton applied to a real case from your firm? Contact us for a demo.