last updated: september 2026

i’m the lead software engineer for applied ai at lexisnexis, where i run technical direction for 20+ production ai workflows in enterprise legal research and the shared llm infrastructure they sit on: retrieval, prompt and version management, model routing.

before that i founded backwork, where i built verity and led a team of 3; and freelanced as an ai engineer for fortune 500 clients, automating etl end to end from scraping to finished product. the things i’ve shipped that i still think about are praetor, a classifier for bankruptcy docket events that beat the legacy svm by 9.5 points and saved ~$25k a week; gavl, a browser agent that navigates court portals to extract case data; and meridian, a scraper that writes and caches its own scrapers across 400+ healthcare sources.

i studied cs at oklahoma state (3.9), after starting at the university of oklahoma on the robotics team.

some fun facts:

• i’ve shipped 63 projects since 2018. some solved a problem. others were interesting enough to build anyway.
• i wrote firmware for a robotic dog and the web app that drives it.
• i trained a forecasting model that beat grid operators in 6 of 7 major u.s. regions with ~40% lower error, then gave it away.
• i ran 45 experiments training models under simulated radiation up to 10,000x leo to see what breaks first.

interests

ai, energy, and robotics are one problem, not three. the limit stopped being algorithms a while ago. it’s power, compute, and heat. a model you can’t power is a demo. a model that can’t move anything is a chatbot.

i’ve trained models, deployed them, written firmware, and built the control software. the physical side is where i keep ending up. energy, machines, the trades. software mostly skipped that part of the economy, and that’s where the work is now.

i’m most interested in systems that run in production, not in a notebook. deployment is what tells you which parts of the research were real. the gap between a benchmark and a thing people depend on is where most of the engineering lives, and it’s underrated.

i also think the tooling under ai agents is worse than it should be. i build a lot of it for myself: toolmux cuts agent tool-schema context up to 96%, sandcastle runs untrusted javascript at 380k ops/sec, nerve validates electrical-harness designs with 43 deterministic checks.

products

• verity searches medicare, medicaid, and commercial payer policies in plain english.
• rebar finds, diagnoses, and fixes breakage in production software.
• faraday handles estimating, scheduling, time tracking, and invoicing for electrical subcontractors.
• tokenbase shows engineering teams how their ai coding agents are used and what they cost.

the rest are on the projects page.

writing and video

on the blog i write about what i’ve built. if the subject is attention mechanisms, i show the math. if it’s deployment, i show the code.

i run applied tensors, where i explain ai concepts and build things live, and ares, where i talk about tech without a script.

the books page lists what i’m reading.

socials

i’m easy to reach. if you’re working on something in ai, energy, or robotics, email me.

email: tylergibbs048@gmail.com
github: tylerbryy
twitter: @tylergibbs
linkedin: tylergibbss
resume: pdf