Approvals are truth
Completion is not enough. The system has to make approval, evidence, fallback paths, and state transitions explicit.
Daniel G. Wilson / Founder / Legion Health
Austin, Texas
I build AI systems that survive contact with reality.
I work on AI control planes, agent UX, and operational software for high-stakes environments. Most of my current thinking comes down to one question: how do you make AI reliable enough to ship when the world is messy, regulated, and full of edge cases?
Legion Health / YC S21Princeton ORFE '18Former Microsoft PMStack Overflow 15K+
01 / Thesis
Model. Harness. Environment.
Reliable AI is not a model-selection problem. The model matters, but production truth is usually decided by the harness around it and the environment it has to survive in.
Verifiability defines throughput
The bottleneck is not code generation. The bottleneck is whether you can tell what is correct, what failed, and what should happen next.
Agent UX is control-plane design
Tool surfaces, safe defaults, ceilings, and operator context all change what an agent can reliably do.
Staged autonomy beats narrative theater
I prefer constrained v1 systems with replayable evidence and explicit gates over big replacement stories that collapse in production.
02 / Selected Work
Founder-led software, operator-grade systems, and long-tail product taste.
The common thread is not a sector. It is software that has to do real work in the world: auditable systems, sharp interfaces, and products clear enough that people can trust them.
2021-now
Legion Health
Co-founder
Building AI-native psychiatric operations with staged autonomy, approval rails, and regulator-grade evidence.
2018-2020Microsoft
Product manager
Learned what coordination debt looks like inside a giant organization and why explicit systems beat alignment theater.
Before and alongsideMeetingbird, Crash Pad, and indie software
Builder
Launches, iOS apps, motion work, and creative tools that sharpened product taste and shipping instincts.
03 / Writing
Working notes on reliable AI.
The writing side of the site is becoming a calmer place for doctrine, field notes, and sharper explanations of what actually makes AI systems work.
AI control planes for messy environments
How approvals, queues, reason codes, and operator surfaces become part of the product.
Model, harness, environment
Why the wrapper around the model usually determines whether an AI system survives contact with production.
AI-readable and AI-mutable systems
What it takes to let agents operate on real workflows without losing traceability, safety, or human control.
The first public essays will focus on model / harness / environment, verifiability, staged autonomy, and agent UX.
Visit /writing04 / Proof
Public traces, kept tight.
Enough proof to establish credibility without turning the homepage into a profile dump.