At Vitreon Legal, we believe the best legal AI is built by people who never stop learning — and we hold ourselves to that standard.
When I started building Vitreon, I made a decision early on: whatever we figure out about making AI work better for legal research, we publish it. Our benchmark results, our methodology, our failures. This isn't altruism — it's how you stay honest. When your results are public, you can't let the bar drift quietly downward.
Learning is embedded in how we build
Legal AI is a fast-moving field. State-of-the-art retrieval models, reasoning benchmarks, hallucination mitigation — these evolve month to month. Our team reads the research, runs experiments, and stays at the frontier because the alternative is building on stale assumptions.
Our GaRAGe retrieval pipeline scores 0.824 RAF — 36% above the published state of the art. Our LEXam Open EN benchmark result is 0.691, surpassing Claude 3.7 Sonnet by 21%. These aren't marketing numbers — they're results from standardised benchmarks you can reproduce. We publish them because we want to be held accountable to them.
Open methodology, not a black box
Too much legal AI is a black box. You get an answer; you don't know why. At Vitreon, we document how our retrieval works, what our training data covers, what our limitations are. The blog at vitreon.app/blog exists precisely for this reason — to share what we learn so the broader legal tech community benefits.
In 2026, we competed in ARLC (Automated Research in Legal Corpus), the first public international legal AI competition. We placed first in warmup with a score of 0.958, and fourth in finals. We published our post-mortem: what went well, what didn't, what we'd do differently. That's the learning culture we operate in — performance is a process, not a destination.
Async work enables real learning
There's a practical side to this commitment. Our team is fully remote and async-first. That means people have time to read, think, and experiment — not just ship tickets. I don't believe you can build thoughtful AI systems on a treadmill of mandatory meetings and reactive work. The flexibility we've built into how we operate is partly what makes a genuine learning culture possible.
What growth means at Vitreon's stage
We're an early-stage startup, which means growth isn't about climbing a career ladder — it's about building frontier competence in a domain that is defining how law will work for the next decade. Every person on the team is both a contributor and a student. We work at the edge of what has been figured out, and we document what we find.
That's the commitment: keep learning, publish what we discover, and build AI that reflects genuine understanding of how legal research actually works.