This is a guest conversation episode of Ship It Weekly, separate from the weekly news recaps.

In this Ship It Conversations episode, I talk with Devansh about why the trickiest legal AI hallucination is a real case used the wrong way, how his team splits their system into three tiers to find failures fast, and why he'd rather optimize for maintenance than for generation.

Devansh is the Founder and Head of AI at Irys. He builds reasoning and verification systems for legal AI, a tough place for hallucinations and confident wrong answers. He also publishes open source AI research on Substack.

We start with why AI for professionals breaks down. Devansh thinks models are getting good at doing a lot of things somewhat well, but they fall short on the last mile to reliable. So his team engineers structure back in instead of hoping a better model fixes it. In legal, fake citations are the failure everyone has heard about. The more interesting one is a real case that doesn't apply. So much of law is interpretation, and a model built to predict the most probable token pulls toward the average.

Then he walks through the three tiers behind Irys One. Ingestion maps messy input, like PDFs with handwritten notes and email chains that refer to them, into a knowledge graph so the AI knows where to look. Reasoning navigates that graph. Delivery applies how a firm or a lawyer likes to work. The payoff is error isolation: if a lawyer says something is missing, two or three targeted queries tell them whether it was ingestion or reasoning. We also get into why chunk-based retrieval misses connections, why going from 10 docs to 10 gigabytes is mostly about finding the few documents that matter, and how "answer contracts" handle "it depends."

Then I take a DevOps angle. Devansh says the hardest part of running AI in production is tracking everything. A fully agentic system might perform 20% better, but you never know how it breaks until it breaks. The first thing his team tracks is whether the system behaves as expected, because labs change models behind the scenes with post-training, quantization and tokenizer changes. We also cover why he calls reinforcement learning the most overhyped technique in AI: it works when the output is truly verifiable, and falls apart when "better" has no real signal.

The big takeaway: you can't fix what you can't find. Engineer for error isolation first, and optimize for maintenance, not generation. Coding agents will churn out heaps of code, but maintenance is where a system lives or dies.

Highlights

• Why a real case applied wrong is worse than a fake citation

• Why AI's pull toward the average struggles with legal judgment

• Ingestion, reasoning and delivery as three separate tiers

• Knowledge graphs vs chunking for finding what matters

• Isolating failures with two or three targeted queries

• Models that change under you, and tracking expected behavior

• Why reinforcement learning is overhyped

• Optimize for maintenance, not generation

Links

Devansh on LinkedIn: https://tsn.io/7Hjfm

Irys: https://tsn.io/ytG1A

Artificial Intelligence Made Simple (Devansh's Substack): https://tsn.io/ZM6eS

More episodes and show notes: https://tsn.io/NqkdP

On Call Brief: https://tsn.io/Gpz2d

LMGT Awards: https://tsn.io/lAERA