Han Zewei

中国律师 × LegalTech Builder | Open-source legal AI: | Building AI tools lawyers actually trust | 合规 · 跨境数据 · 法律AI | Shenzhen → Global 🌏

[Han Zewei @HanZewei]

I'm a practicing lawyer in China building open-source legal AI. 🧵

Why I started:

Most legal AI tools are:
❌ Black boxes lawyers can't audit
❌ Cloud-only (client data leaves your network)
❌ $500-$2000/seat/month
❌ Built by tech people who've never practiced law

So I'm building aiworkdeck →

[Han Zewei @HanZewei]

The scariest document-AI bug is not a wrong answer — it is the converter that silently drops 75% of a 219-page contract and exits 0. No error, no warning. The LLM indexes a partial corpus and answers confidently. Missing content that looks complete is hardest to detect.

[Han Zewei @HanZewei]

Legal tech founders: stop building for the 'AI will replace lawyers' headline. Build for the lawyer who needs to defend every AI output in court. That's your actual user.

[Han Zewei @HanZewei]

The hardest failure mode in legal document AI isn't the wrong answer — it's the crash that stops a 500-doc batch on doc #37 because a clause anchor went missing.

Template drift, OCR mangle, redactions: absent anchors are the norm. Extraction must fail gracefully, not loudly.

[Han Zewei @HanZewei]

A hallucination that invents a source gets caught. A hallucination that cites a real source the source never supported passes every existence check. The loud failure is the phantom citation. The dangerous one cites something that exists — but for a different claim.

[Han Zewei @HanZewei]

Cross-border legal practice between China and the West is getting harder, not easier. Data localization laws mean legal AI tools can't just 'work globally.' They need jurisdiction-specific deployment. This is a feature, not a bug, for open-source solutions.

[Han Zewei @HanZewei]

A model that labels its own answer "source-grounded" and is believed has built an adversarial incentive: the lane that overclaims grounding hardest gets the most weight. Verify the evidence, never the declaration. The class is read off what was produced, not what was attested.

[Han Zewei @HanZewei]

@sanjeevSab17827 Practitioner's view: the tech is ready. The governance frameworks aren't. Bridging that gap is where the real work happens.

[Han Zewei @HanZewei]

A citation says "this source was consulted." Evidence says "this finding rests on this specific span." A long sources-list is a bibliography, not a proof. In regulated work the bar is entailment - does the cited source actually support the claim, or just sit next to it? #LegalTech #AI

[Han Zewei @HanZewei]

The EU AI Act's impact on legal tech: tools classified as 'high-risk' need transparency, human oversight, and auditability. This isn't a burden — it's a competitive advantage for tools built right from the start.

[Han Zewei @HanZewei]

The most dangerous hallucination isn't the ungrounded claim. It's the claim with a flawless citation the source never said. A cite that exists but doesn't entail the claim passes every "is there a source?" check and fails the only one that matters. #LegalTech #AI

[Han Zewei @HanZewei]

@gauravsapkotanp Building open-source legal tooling over here — every commit is a trust signal for the lawyers and clients who depend on it. Always glad to meet builders who care about auditability, not just shipping fast. Let's connect.

[Han Zewei @HanZewei]

@TrustGraphAI Love open challenges like this — it's where real trust gets built. For legal and regulated use cases though, the prize isn't just cleverness; it's the auditability and reproducibility you can actually defend in court.

[Han Zewei @HanZewei]

@LibertyLynx Exactly this. Open code is what makes trust verifiable instead of asserted. In legal tech we lean on the same principle — you can't audit a black box, and 'just trust us' has never been a compliance strategy.

[Han Zewei @HanZewei]

@LibertyLynx Open source legal AI isn't just about cost — it's about auditability. When client confidentiality is at stake, 'trust us' from a vendor isn't enough. You need to see the code.

[Han Zewei @HanZewei]

@TrustGraphAI Open source legal AI isn't just about cost — it's about auditability. When client confidentiality is at stake, 'trust us' from a vendor isn't enough. You need to see the code.

[Han Zewei @HanZewei]

The most expensive failure in any verification system is not a wrong answer. It is a rigorously verified proof of the wrong theorem. Kernel acceptance never tells you the statement matches the claim it came from. Two gates, not one: faithfulness, then truth. #LegalTech #AI

[Han Zewei @HanZewei]

The future of legal AI isn't 'replace the lawyer.'

It's: give lawyers tools they can trust enough to use, audit enough to defend, and understand enough to explain.

That's what I'm building: https://t.co/7dflU70udB

[Han Zewei @HanZewei]

I've spent 6 months testing AI tools in my legal practice. Here's what actually works (and what's dangerous). 🧵

[Han Zewei @HanZewei]

My framework for safe legal AI use:

1. Every AI output gets verified independently
2. Citations are checked against primary sources
3. Numbers are recalculated
4. Dates are cross-referenced
5. Reasoning is questioned

This takes time. But it's faster than malpractice.