Legal AI development — citation-grounded research, contract intelligence, intake automation.

Five layers. Hybrid retrieval grounds every answer in your authoritative corpus (case law database, internal brief bank). Required citations link every legal claim to a retrieved source — answers without citations are rejected before display. A refusal layer activates when authority is missing or out of jurisdiction. Confidence scoring routes low-confidence answers to attorney review. An eval harness blocks any prompt or model change that regresses citation accuracy against the partner-reviewed golden set.

Yes. Mutual NDA is signed before any technical or matter detail is shared. For engagements involving privileged client communications or work product, we also sign confidentiality addenda that match your firm's standard outside-vendor terms. Self-hosted Llama 3 inside your VPC means client data never leaves your infrastructure — we operate the engineering, you operate the data perimeter.

No, and we will not build it that way. Every Aiinfox legal AI deployment is built so the AI accelerates the attorney (research, drafting, document review, intake) but the attorney signs off on every output that leaves the firm. The refusal layer is scoped to be aggressive on legal-advice queries — the AI will not give legal advice to end clients, only assist authorised attorneys. This is both an ethical / regulatory requirement and the only design firms actually adopt at the partner level.

Jurisdiction-agnostic at the platform level — we have shipped legal AI for US (federal + state), UK, EU (jurisdiction-aware retrieval), and Indian (Supreme Court + High Court) legal corpora. Each engagement requires a jurisdiction-specific eval set + citation format + statutory framework configuration. We work with your firm's research librarians and senior attorneys to scope the jurisdiction correctly.

Six to eight weeks for a focused v1 — a research agent for a specific practice area, a contract intelligence pipeline for one document type, or a legal intake chatbot. Twelve weeks for a firm-wide knowledge-base RAG over brief banks + memos. Larger e-discovery deployments are typically 14-20 weeks including secure document ingestion and reviewer workflow integration.

Most legal AI v1 engagements at Aiinfox land between $40,000 and $150,000 fixed-price depending on jurisdiction scope, corpus size, and integration complexity. Firm-wide knowledge-base RAG over a multi-decade brief bank typically reaches $120,000 to $250,000. Per-matter document review and e-discovery deployments are typically priced per-document or per-matter rather than fixed v1.

Yes. For privileged client data, we deploy Llama 3 70B or 8B on vLLM inside your firm's AWS / Azure / GCP VPC, or on-prem on firm-owned hardware for the highest-sensitivity matters. Self-hosted vector store (pgvector / Qdrant) keeps retrieval local. Zero data egress. Audit logs are exportable for partner review and ethics-committee documentation.

Permission-aware retrieval — each user's access scope (which matters, which clients) is part of the query and filters the vector store at the index level before similarity search. The AI cannot retrieve documents from a matter the user is not authorised to see, regardless of how the query is phrased. Audit logs record every access attempt for ethics-committee review. We integrate with existing conflict-check systems where they exist.