Eric Zayler

I’ve made and lost real money letting code trade for me, which is why I care more than most about systems that fail safely. I build autonomous AI agents in Python, ship production AI software, and do honest, reproducible ML research. I’m mostly self-taught — I left a computer-science and cybersecurity path to build and trade full-time, and nearly everything here I learned by shipping it. What I’m after now is where AI value lands, and how to keep these systems safe as they get more capable.

Writing

Discipline belongs in the code, not your willpower What being force-liquidated by my own override taught me about building constraints that hold when you won’t — and why it’s the same problem as AI safety.

Work

bert-meaning-probe A from-scratch reproduction of Chronis & Erk (2020): which layer of BERT best captures word meaning, and why depth only helps once you give the model context. Honest scope — a chance floor, a permutation test, a reported null, and CI.
Vexara A production voice + SMS AI platform I built and operate — Twilio, Vapi voice AI, a browser softphone, 118+ passing tests on a live VPS.
Autonomous trading agents private Kelly-fraction position sizing, regime detection, automated kill-switches, and an investment policy enforced in code rather than willpower — with breakers that revert the system to simulation before I can override them.
Agent infrastructure & safety tooling private A fleet of OAuth-gated MCP servers behind a hook-based layer that gates dangerous operations before an autonomous agent can run them. Found and fixed a live authorization-bypass in my own infrastructure.

Spearman correlation against human similarity by BERT layer: context-free word vectors decay with depth, while multi-prototype embeddings with MaxSim rise to a middle-layer peak around layer 8. — From bert-meaning-probe: a single context-free vector per word loses the signal with depth (grey); give each word real usages and cluster them per layer, and the curve inverts into a middle-layer peak (blue). The depth was always doing useful work — you just need context to use it.