About
I've been taking systems apart since I was nine.
First games. Now legal decisions and aviation risk. Same instinct, much higher stakes.
I'm Dylan Taylor — 17, in the San Francisco Bay Area. I taught myself to program by reverse-engineering video games: from about nine, I pulled game binaries apart to understand how they kept their world in memory, and I built cheats — first for Fortnite, then Valorant. It grew into a real operation with actual users, run for a couple of years against professional anti-cheat teams whose entire job was to shut me down.
Then Riot's lawyers sent a cease-and-desist, and I stopped. I was a kid; I owned it; I moved on. But that era is where I actually learned to build — to teach myself hard things with no map, ship under pressure, and think in adversarial systems. I took the same instinct — take a system apart, find where it breaks, build something that holds — and pointed it at problems that matter. The whole arc is here.
Now I build AI for high-stakes decisions — the places where a fluent answer isn't enough and you need one you can defend, weeks later, to someone who wasn't in the room. In law, I'm building Littman, which turns a firm's own standards into structured, cited decisions. In aviation, I built FlightReady, a risk prototype that treats a go/no-go call as a bet on the worst plausible day, not the average one. Different industries, one shape: make the judgment legible, and keep a human in the loop.
Finish it fully.
I don't switch off a problem until it's actually done. If something breaks at 2am, I'm up until it's fixed, tested, and I can explain why it broke — not just why it works now. "Done" means merged, tested, and I could walk someone smarter than me through every line. A half-finished feature isn't progress; it's debt I'll pay back at a worse hour.
Most of my best work happens between midnight and five, when nothing pings and the problem is the only thing in the room. It's less discipline than not being able to leave it alone. When something's too big to hold in my head, I break it down until each piece is obviously true or obviously false, then start with the piece I understand least. And I assume the bug is my fault until the evidence forces me elsewhere — it usually is, and finding out exactly why is the fastest way to stop repeating it.
Train the thing that compounds.
I lift most mornings before I open a laptop. The gym is where I practice doing a hard thing on a day I don't feel like it — which turns out to be the exact skill a stuck bug demands at 3am. Cold exposure and long runs aren't a personality; they're cheap, repeatable reps at being uncomfortable, so discomfort stops registering as a reason to stop. I train the body for the same reason I test the codebase: it compounds, and skipped sets and skipped tests both come due later, with interest.
What I reach for
Not a proficiency chart — the layers I actually work across.
- Retrieval & RAG
- Hybrid search, reranking, citation grounding, confidence and abstention.
- Decision engines
- Structured extraction, neuro-symbolic rules, audit traces, human-in-the-loop.
- Risk modeling
- Weighted, probabilistic, and tail-risk (CVaR) methods — and knowing when each lies.
- Adversarial systems
- Reverse engineering, red-teaming, robustness — building things that fail loudly.
- Product engineering
- React, TypeScript, Next.js, Supabase, Python — shipped end to end.
- Writing & explanation
- Turning a complicated idea into something you can hold in under a minute.
When I'm not building
- Powerlifting
- chasing a 2× bodyweight deadlift, tracked honestly
- Long early runs
- before the day starts pinging
- Blitz chess
- and too much endgame study
- Cold exposure
- cheap reps at being uncomfortable
- Mechanical keyboards
- built, not bought
- Reading on paper
- systems papers and builder biographies
This site
Designed and built from scratch — no template, no page builder. Next.js and TypeScript in strict mode, Tailwind, statically generated. Six interactive labs, each on a pure, unit-tested engine. Built to WCAG 2.2 AA: keyboard-navigable, reduced-motion aware, readable from a 320px phone up. The look is "flight recorder" — warm ink, instrument amber, monospace telemetry.
It was also built the way I build now: a team of specialized AI agents doing bounded, parallel work against tight specs, with an independent verifier catching what any single confident pass would miss. You can step through that exact process in the build orchestrator.