Dr.Legal enterprise services

On-premises legal AI

A legal large language model and Taiwanese legal retrieval system, deployed in your own server room, so that question answering never leaves your network. With a quantified hallucination acceptance method and ISO 42001 governance alignment.

Three things that block legal teams from adopting AI

Case files and contracts cannot leave the network

For finance, healthcare, the public sector and trade-secret matters, sending documents to an external AI service is itself a compliance problem. On-premises deployment keeps the data in your own server room.

General-purpose AI cannot find Taiwanese law

A model with no connection to a Taiwanese legal database risks fabricating statutes and judgments. Legal teams cannot rely on it, and verification ends up costing more than it saves.

Nobody can say how reliable the AI is

The hardest step in adopting AI is explaining its reliability to management and auditors. Without quantified evidence, projects stall at the pilot stage.

What we deliver

On-premises legal LLM deployment

A legal large language model and retrieval system deployed as a whole on your own GPU servers. Question answering runs entirely inside your server room; nothing is sent out. Configuration follows your existing hardware and concurrency needs.

Taiwanese legal corpus and retrieval

Retrieval over Judicial Yuan judgments, statutes, administrative rulings and Constitutional Court decisions. Answers retrieve first and generate second, with every source cited back to the original text.

Hallucination audit and acceptance testing

Model faithfulness is measured before and after deployment by double-blind human audit, producing quantified metrics you can write into acceptance clauses, not just a subjective trial impression.

ISO 42001 governance alignment

AI decision audit logs, risk event detection, model performance monitoring and compliance reporting, aligned with the ISO 42001 AI management system for internal and external audit.

Integration with existing systems

Standard interfaces into your document management, case management or knowledge base, plus MCP so your internal AI assistants can call legal retrieval directly.

Calibrated by practicing lawyers

A practicing lawyer takes part in design and acceptance, so output is judged against the standards of legal practice, not only from an engineering point of view.

How we prove the AI can be trusted

The hardest part of adopting legal AI is usually not whether the model is useful but how to explain its reliability to management and auditors. We treat reliability as something to be measured. The method is below, and all of it can be re-run inside your environment.

  1. 1

    Define the categories

    Break 'hallucination' into separately measurable classes: fabricated reasoning, quotations that differ from the source, citations of judgments that do not exist or were never read, and conclusions contrary to the cited decision. Each has its own denominator; they are not blended into one number.

  2. 2

    Sample real traffic

    The test population is a stratified sample of questions from real usage, not questions we wrote ourselves. Self-written test sets systematically overstate performance.

  3. 3

    Double-blind human audit

    Several auditors judge independently; agreement rates are computed and disagreements go to arbitration. The agreement rate is part of the report, so the credibility of the numbers can itself be checked.

  4. 4

    Mechanical re-check

    Whether a cited case number actually exists, and whether it was within the material the model read, is verified programmatically, item by item, not by human reading.

  5. 5

    Frozen inputs, reproducible results

    Evaluation inputs are packaged and hashed. Anyone can re-run the evaluation from the same inputs and get the same result, independent of time or environment.

The methodology itself is one of the deliverables. You can apply the same method to evaluate other vendors’ AI systems, or adopt it as the acceptance standard in your internal AI governance.

Who it is for

  • Finance and insurance (data residency and regulator requirements)
  • Technology manufacturing (trade secrets and patent litigation files)
  • Public sector and state-owned enterprises (procurement and security rules)
  • Healthcare (medical records and personal data protection)
  • Law firms (client confidentiality obligations)

Proven in production

The same technology runs the free public platformDr.Legal, covering 22.8 million judgments plus statutes, administrative rulings and Constitutional Court decisions, and handling a substantial volume of real legal questions every month. The retrieval tool’s source is published onGitHub, so you can inspect the quality of the implementation before deciding to talk to us.

Frequently asked questions

How is on-premises different from using a cloud AI service?

The difference is where the data goes. On-premises, questions and documents stay inside your own server room and never pass through an external provider. For organizations under regulatory supervision, with trade-secret concerns or confidentiality duties, that is usually a question of feasibility, not preference. Note that the public Dr.Legal website generates its AI answers through an overseas cloud model provider and is not on-premises; the on-premises environment is built separately under this enterprise service.

What hardware do we need?

A server with GPUs. The actual specification depends on the number of users, concurrency and response-time requirements. The assessment phase first checks whether your existing hardware is sufficient, to avoid unnecessary purchases.

How do you prove the answers are reliable?

By double-blind human audit combined with mechanical re-checks, producing quantified metrics for your acceptance clauses together with the full method and auditor agreement rates, so the numbers themselves can be examined. The same methodology can be used to evaluate your existing AI systems or those of other vendors.

Do we have to hand over our data?

No. Your case files and contracts stay on your side, and evaluation during deployment can be carried out inside your environment.

Can we start with a small pilot?

We recommend it. A limited proof of concept first confirms integration and results; the deployment is widened once that is established.

How does this relate to the free Dr.Legal platform?

It is the same technology. Dr.Legal is the public, free platform and also the proving ground for this technology, handling a substantial volume of real legal questions every month. The enterprise line is delivered personally by Aaron Huang, Attorney at Law, who also signs the contracts.

Tell us about your situation

You do not need a specification yet. Describe the current state and the constraints; the assessment phase works out what is feasible together with you.