Leta — Legal AI Assistant

Law, Expertise, Transparency, Access — for lawyers, journalists and corporate legal teams.

The problem

Lawyers spend hours searching the Official Gazette and case law. Generic GPTs hallucinate non-existent articles. Cloud solutions break attorney-client confidentiality.

How it works

1
We index the Romanian Official Gazette (1990→present), legislatie.just.ro, ICCJ rulings and CNSC decisions — over 2.8 million documents, refreshed daily.
2
Ask in natural Romanian. The system performs hybrid retrieval (semantic + keyword) over the real corpus — not over generic 'world knowledge'.
3
Answers include links to original sources for each claim. Our Anti-Hallucination pipeline blocks the answer if it can't cite a source.
4
On-premise deployment in 2 weeks for sensitive practices. Data stays on the client's infrastructure — zero cloud, zero training on your corpus.

Capabilities

Anti-Hallucination v4

Source-validation pipeline on every answer. If no reference exists in the corpus, the system says 'I don't know' — never inventing.

Citations linked to the original source

Every legal claim links to the exact article in the Official Gazette or ICCJ ruling. Auditing becomes 30 seconds, not 30 minutes.

Daily automatic indexing

Newly published Official Gazette acts enter the corpus within 24 hours. Answers reflect the current state of law, not last year's snapshot.

On-premise for sensitive data

For practices working in M&A, criminal or insolvency, we provide dedicated deployment within the client's infrastructure. No data leaves your network.

Multi-corpus: legislation + case law + practice

Complex legal questions combine code, relevant rulings and doctrinal opinion. The system searches all three in parallel and synthesizes transparently.

Web app + Android

Access from anywhere — desktop for deep research, phone for quick questions in meetings or before the bench.

Tech stack

▸ Romanian-language LLM fine-tuned
▸ Hybrid retrieval (BM25 + dense vectors)
▸ Citation-grounded generation
▸ On-premise / private-cloud EU-resident

Evidence

✓ 2.8M+ Romanian legal documents indexed
✓ Sub-3-second response for factual queries
✓ Guaranteed on-premise deployment in 2 weeks
✓ Audit-ready citation log for every answer

FAQ

How do you guarantee the AI doesn't hallucinate non-existent law articles? +

The Anti-Hallucination v4 pipeline requires every factual claim to have an explicit reference in the corpus. If the LLM cannot produce a verifiable citation, the answer is blocked and the user sees 'insufficient information in corpus'.

Does our case data enter any public corpus? +

Never. In on-premise deployment, prompts and answers never leave the client's infrastructure. The model is not retrained on your corpus — it uses RAG, not fine-tuning, precisely for this reason.

How long does on-premise implementation take? +

Two weeks standard: week 1 — provisioning + ingestion of client corpus (if any); week 2 — integration with client systems (Active Directory for SSO, document management) + team training. Cost includes 12 months of support.

How does it compare to ChatGPT, Claude or other general LLMs? +

General LLMs were trained on the public internet, where the Romanian legal corpus represents <0.3%. They answer fluently but invent law articles for specific questions. Leta searches a dedicated corpus of 2.8M RO documents and cites the exact source.

Is there a trial? +

Yes — the public version on lege365.ro allows limited free questions. For law firms, we offer 14 days of full access with no question limit, after a 30-minute conversation about your needs.

Does it support European case law (CJEU, ECHR)? +

Currently the base corpus is Romanian + directly-applicable EU regulations (GDPR, AI Act, etc.). CJEU and ECHR case law are on the H1 2027 roadmap — for interested enterprise clients, we can ingest additional private sources now.

We start with a 30-minute conversation.

Free AI-readiness audit for companies with 50+ employees. We reply within 24 hours.

Open contact form →