Framework
The Three Categories of Hallucination
Devansh breaks legal AI hallucinations into three categories. The third is the most dangerous, and the hardest to catch with traditional vector search.
How it works
- 1Citation hallucinations: The AI cites a case that does not exist.
- 2Applicability hallucinations: The case exists, but the jurisdiction, domain, or context makes it inapplicable.
- 3Context hallucinations: The AI misses a relationship between documents, where one document modifies, contradicts, or conditionally applies to another. Irys addresses this with a self-updating knowledge graph that links entities, propositions, and assertions across the entire document set.
Hear it on the podcast
Legal AI: Why Lawyers Are Finally Free to ThinkListen to the episodeMore from this episode
The Fragmentation ProblemMost legal AI today is what Devansh calls "a system prompt wearing a trench coat", a niche product wrapping a general-purpose model, charging per word or per page for the privilege. Small and mid-sized law firms end up stitching together 10 point solutions that share no context with each other. Irys attacks this from the foundation.The Democratization MissionDevansh grew up watching legal inaccessibility cause real harm, civil cases in India carry a 10-year backlog; New York City tenants get bullied by landlords because they cannot afford to fight. His co-founder, a former Big Law attorney, had lived the inefficiency from the inside.
Get the next framework first
One email a week with the frameworks, tools, and founder stories from the podcast. Free.
More frameworks for AI & Automation
The Round Table of Second OpinionsInstead of marrying one model, poll all of them. Each model has different training data, different personality, different blind spots.The Unidentified Data PrincipleMost apps say encrypted, in transit, at rest, de-identified. Stephen goes one step further with an architecture where the business model physically cannot pivot into data harvesting.The LLM-as-Wireframe MethodDavid's practical framework for using AI responsibly in legal contexts without replacing professional counsel entirely.Deterministic Logic Plus Conversational AIMost online will builders are wizard-based forms. You fill in fields, answer dropdowns, and a document gets generated. The problem is that approach assumes you already know what you want. That's almost never true. Succession Wills solves this with a split architecture.Personalization as the North StarThe current AI DJ surfaces the same highlights for all users. The roadmap moves toward a fully personalized agent.AI Disqualification StackUse AI to kill bad deals before they reach human underwriting. Build a hardcoded list of non-negotiables and run every inbound deal through it as a first pass. The goal is not to automate the investment decision - it is to automate the no. Five deals worth real attention each week cannot receive that attention if the team is wading through fifty non-starters. The edge is focus, not volume.