Five hard problems in legaltech AI

1. Cross-clause reasoning
   Most contract review tools classify clauses in isolation. They flag a one-sided indemnity, a missing notice period, an aggressive non-compete. Fine. The harder problem is when two clauses are individually reasonable and dangerous together. A modest liability cap is normal. An "all damages including consequential" clause is also normal. Put them in the same contract and the cap is meaningless because the consequential damages route around it. Embedding-based retrieval doesn't catch this. You need a graph representation of clause dependencies, or a second pass that reasons over the full document with explicit relational prompts.

2. The training data isn't Indian
   Almost every legal LLM benchmark — CUAD, LexGLUE, ContractNLI — is built on US or EU contracts. Models trained or fine-tuned on these miss things that matter in Indian contracts. Stamp duty implications. State-specific labour law. The Indian Contract Act's treatment of consideration. The Specific Relief Act on remedies. If you're building for India, you'll either fine-tune on Indian contracts (which are mostly not in any public dataset), use retrieval-augmented generation with Indian statutes, or accept that your tool will be right on common clauses and confidently wrong on Indian-specific ones. There's a real opportunity in just curating clean Indian legal datasets, separate from any model work.

3. Redlining is structural, not generative
   A redlined contract isn't a new contract. It's a diff — additions, deletions, and modifications anchored to the original text. Word's Track Changes is a structured XML representation, not free text. Most LLM-based contract tools output a "suggested rewrite" as plain text. That's a draft, not a redline. Getting from suggestion to actual redline requires diff-aware generation — either constraining the model to produce structured edit operations, or post-processing the output through a diff algorithm that can survive the small rephrasings LLMs love to add. Both approaches break in interesting ways.

4. Plain English isn't translation
   "Indemnify, defend, and hold harmless" is easy to translate. "You'll cover their losses" is fine. The hard part is telling a specific user why this clause matters for them. A two-person SaaS startup signing with an enterprise client cares about indemnification differently than a freelancer signing a one-off NDA. Same clause, same translation, completely different significance. The interesting builds combine clause classification with a lightweight user profile (industry, transaction value, role) to filter what's worth flagging.

5. Regulatory data goes stale weekly
   If you're building anything in the legal governance track, your knowledge cutoff is your biggest enemy. The Companies Act gets amended. SEBI puts out a circular. The Ministry of Labour updates a notification. A model trained six months ago doesn't know any of this.
   RAG over a regularly-updated source corpus is the cleanest answer, but the harder problem is freshness of the corpus itself. Most Indian regulatory data is behind PDF circulars on different government sites. There's no API for "what changed this week in Indian labour law." Building one is, itself, a hackathon-worthy problem.

The hackathon

If any of the above sounds like a weekend you want, here's where to do it.

The Code of Law Challenge. 06–07 June 2026.

Pick one of two problem statements:

1. Contract Review & Redlining Tool. Reads an NDA, employment, or service contract the way a careful lawyer would. LLMs, rules, hybrids — your call.

2. AI for Legal Governance (open track). Anything that automates compliance, policy review, regulatory tracking, or legal workflows using open-source tools.

What you get

• ₹22,000 prize pool, split between the top two problems

• Paid internship offers at Rhett

• Incubation access through LIIC at HPNLU

• One-year ISAIL membership for top performers

• A LegalWiki workshop and certification

Why this hackathon matters

The problems above aren't theoretical. They're what we solve at Rhett every day, building AI legal infrastructure for India's MSMEs. The Code of Law Challenge brings together India's leading legaltech ecosystem — ISAIL (Indian Society of Artificial Intelligence and Law), LegalWiki, and LIIC at HPNLU — to back the builders working on them.

Cash prizes. Direct internship offers at a company actively hiring. Incubation access through HPNLU. A network that opens doors in a space where opportunities are usually closed off to outsiders.

Indian legaltech is at an inflection point. Half the market thinks AI replaces lawyers in two years. The other half thinks AI can't do anything useful in law. The truth is in the middle, and the people who figure it out get to build the next generation of legal infrastructure here.

That's what this hackathon is for.

Early bird closes 21 May. Applications close 04 June 2026.

Register: Google Form Hackathon Website: rhett.legal/hackathon Hackathon Details: brochure