DOKISTDOKISTDOKIST builds AI systems that read your contracts and filings — and answer questions instantly, with the exact source attached.
§4.2Either party may terminate this Agreement upon thirty (30) days written notice if the other party materially breaches any provision and fails to cure such breach within the notice period.
Confidential Information shall remain protected for five (5) years following termination.
✓ Verified — MSA_2024.pdf, p.12 ⚠ Cure period shorter than your standard termA search assistant built over your contracts and filings. Ask a question in plain language, get an answer with the exact source cited — no more digging through hundreds of pages.
AI that reads a document, researches the context, and drafts the risk memo or due diligence report your team would otherwise spend hours writing by hand.
Every system is tested and tuned on your specific contracts and language — not a generic template — so it understands your documents, not just documents in general.
What each one does, in plain terms — and the measurement behind it.
Reviews a full document set and drafts the due diligence report. Cut unsupported answers by more than half against its own first version, measured on a fixed benchmark set. About 4 seconds end to end.
Same question set, before and after; both configurations versioned, neither regenerated at read time. Roughly one claim in eight still lacks a clean span binding — flagged in the memo, never dropped. A system reporting 0% at this sample size would be reporting its own blind spot.
Answers questions across a whole contract library with every answer linked to its exact clause and page, and flags unusual or risky clauses so reviewers only look at what needs a second read. Retrieval is tuned on legal language specifically, so clauses that read alike but mean opposite things stop colliding.
The load‑bearing null result: a rank‑8 LoRA on the generation side scored 0.81 faithfulness against plain RAG's 0.84 at n = 40 — inside the noise floor. RAG shipped at a quarter of the inference cost, and the work moved to retrieval, where the number above then moved.
Assembles a structured briefing from raw sources in seconds. Traced end to end, so every step of the agent’s reasoning is inspectable.
The remaining 6% is upstream rate limits and transient source failures — handled by degrading to a partial brief with a visible gap list, not eliminated. p95 is reported rather than the mean: a mean of 0.9 s hides a p99 of 14 s, and the p99 is what a user calls broken.
3 reference systems across legal and financial documents — each measured before and after tuning, on public corpora. No client deployments yet.
See the code and the numbersEnterprise players build horizontal tools for Fortune 500 budgets. This studio builds only for legal and financial documents — including the French- and Arabic-speaking, civil-law markets they don't serve.
On French and Arabic — where we actually are. Every number on this page is measured on English corpora. The French and Arabic work is a documented twelve-session research programme, not a shipped capability: a text-layer inventory of the Bulletin Officiel archive, a five-system Arabic OCR comparison on hand-verified pages, and a public French/Arabic legal retrieval benchmark verified by a jurist. No such benchmark exists today. When it does, the numbers replace this paragraph.
Tell us about your documents and what you need to find in them. We'll suggest the right starting point — no jargon required.
Your documents stay yours. Before any work begins we agree in writing where they are stored, which models see them, and who has access — NDA as standard, self-hosted deployment on request.