Chicago, IL / Open to work

Bazyl
Horsey

Serial co-founder and two-time VC-backed startup CTO. Applied AI: agentic systems and edge inference.

I build AI that has to work in production, not in a notebook.

An example day-0 napkin: the customer's site feeds edge inference, which feeds the platform, which feeds agents and the people who approve their work. Misses loop back into retraining.FIG. 1 / A DAY-0 NAPKIN, WRAPPED IN CSSTHEIR SITEtheir systems, data and ITraw dataEDGE INFERENCEx86 · ARM · TensorRT · 26 msevents, resultsPLATFORMmodel registry, pinned per customertasks, alertsAGENTS + PEOPLEagents draft, a human approvesmisses → retrain → re-pin

01 / the claim

Most AI dies in procurement and uptime, not in accuracy.

  1. Ship the 0→1, then make it survive.

    A platform is a demo on day one and an uptime problem on day ninety. I have taken one from the first commit to about twenty customer deployments, which is a different job than the one people pitch.

  2. Replace the API call when the API call is the bottleneck.

    GPT-dependent detections moved to in-house models, and one feature's spend dropped 6×. A new segmentation model took inference from 400 ms to 26 ms and GPU load at five stations from 100% to about 30%. Same accuracy requirement, different bill.

  3. Own the boring parts.

    ISO 27001:2022 from a standing start. SOC 2 Type I readiness. RLS multi-tenant Postgres. Hardened Linux on x86, micro-x86 and ARM. Three-minute median first response. Upkeep is hard, and it is also most of the job.

  4. Grow the people, then get out of the way.

    As the team went from solo to nine across time zones, my week went from 80% code to 20%. The job became getting other people to their best work, and owning my mistakes out loud so they could own theirs.

said about working with me

  • Bazyl has a habit of making complicated things look suspiciously easy. The kind of person you want around when things break, ideas get weird, or someone says, “That’s probably impossible.” He’ll either find a solution or make the problem regret showing up.

    Sherif OkunadeWeb Developer

  • Passionate technology leader who is constantly geeking out over the newest tech releases and figuring out how they can solve real problems.

    Matthew BronsteinHead of Engineering

  • A natural mentor who empowers engineers. I started exploring systems programming and Rust because of an impressive project I saw on his GitHub.

    Jesse OnyekweluWeb Developer

  • His depth of knowledge, resourcefulness, and unwavering commitment to excellence make him … someone you can always count on.

    Oshoke OyatiWeb Development Lead

02 / work

Things that exist and run.

  • Jorn

    elmv0.0current

    Write your UI once in Elm and decide at deploy time where each piece runs: browser, server, or both. Pure Elm virtual DOM, a diff that runs anywhere, a JSON wire protocol and a dependency-free browser patcher. No kernel code, no fork. This page runs on it.

    1Pure Elm vdom, diff, wire protocol, event routingdone
    2Browser patcher, applies patches to the real DOMdone
    3Full HTML/SVG/MathML, keyed lists, effects, runtimenext
    4Server runtime over a websocketplanned
    5Generated, compiler-checked encoders for shared typesplanned
    6Session snapshots and typed migrations across deploysplanned
    https://github.com/bazylhorsey/jorn
  • obsidian-mcp-server

    typescript22 stars

    An MCP server that connects local and remote Obsidian vaults to desktop AI tools: multi-vault switching, graph analysis, hybrid search and backlinks.

    https://github.com/bazylhorsey/obsidian-mcp-server
  • prompt-rust

    rustupstream

    A proposal to add typed stdin prompts to the Rust standard library (rust-lang/rust#117852, RFC 3196). What upstream actually feels like from the outside.

    https://github.com/bazylhorsey/prompt-rust
  • Aethos

    rust

    A Rust take on Phoenix LiveView, built to learn the browser-versus-server trade-off from the implementer's side. The only way to have an opinion about it.

  • Lustre

    gleamupstream

    Hayleigh Thompson's Gleam web framework. Four open PRs, including a 337-line fix for server component context subscriptions and a CI workflow for main. Upkeep is hard. Apparently I do it anyway.

    https://github.com/lustre-labs/lustre/pulls?q=is%3Apr+author%3Abazylhorsey
  • Frigate

    pythonupstream

    The open-source NVR. A merged fix to its auth cookie pattern and an open one for a shared-memory leak on skipped frames. Same class of bug as the day job, in someone else's codebase.

    https://github.com/blakeblackshear/frigate/pulls?q=is%3Apr+author%3Abazylhorsey
  • astonish

    python

    A ChatGPT 3.0-era Lovable, presented at the OpenAI Hackathon in Chicago, 2022. Kept because it is funny and it was early.

    https://github.com/bazylhorsey/astonish

03 / path

Where the scar tissue came from.

  1. 2023→nowChicago, IL

    ArvistCo-founder & CTO, now Technical Advisor

    Techstars '22, then a $4M seed led by Rackhouse. Took the platform from first commit to about twenty customer deployments across five product lines, then built the plug-and-play inspection product behind the pivot: installs in a day, works with the cameras and systems a site already has.

    Grew engineering from solo to nine across time zones. As CISO, built the ISMS from nothing to ISO 27001:2022 and SOC 2 Type I readiness. Principal investigator on a $1M+ DoD-funded MxD safety AI project.

    20customer deployments, first commit to running on site
    9engineers on the team, grown from solo
    131versioned releases in a single quarter
    10.9→3days, median time to close a bug
    90%+of customer incidents traced to site power, network or hardware, not our code
    3 minmedian first response to a customer
    400→26ms per inference after a model swap
    6×less GPT spend on one feature, same output
  2. 2019→2022Chicago, IL

    KoloniSoftware Manager → Application Manager → CTO

    Smart-locker and mobility IoT, now Accessa, backed by Jason Calacanis with a $1M pre-seed. Part-time through college, CTO from 2021, leading a team of five. Built the BLE SDKs and mobile integrations behind the 2020 Smart Locker launch, which became the core business.

    Led the hardware too: our own System on Module with BLE and LTE for the lockers, and the APIs that wrapped it. The software was the easy half.

  3. 2020→2021Storm Lake, IA

    LevityCo-founder & CEO

    First startup out of Buena Vista's Foundry incubator. Wound it down in 2021 and brought two of its engineers to Koloni, which I count as the best outcome any of my companies had. Earlier: the app for Match-Learn-Launch, a $175,000 Google.org Impact Challenge winner.

  4. 2016→2020Storm Lake, IA

    Buena Vista UniversityBS Computer Science, minor in Communications

    Dean's List. President of the Robotics Club. Collegiate wrestler, which explains more about how I run engineering than any conference talk does.

04 / writing & talking

Said out loud, on the record.

Research

peer reviewed / 2024 / second author

Implementing concepts from green logistics in the turkey production supply chain

Frontiers in Sustainability 5:1416964. Modeled roughly 50% lower emissions by rerouting livestock between brooders and finishers with a Gurobi optimization over distance, exclusivity, and capacity. The notebook is public, which is the point of publishing.

Podcasts

software leaders uncensored / with steve taplin

Doesn't sugarcoat it

Academic burnout through to running engineering at Arvist. Why scrappy beats flashy when you are building a team, and what most CTOs get wrong about leadership. The episode title is the whole review.

youtu.be/mk7CKtatU8U

deep learning leaders / with eric landau

Building Arvist, and what AI actually needs in a warehouse

Encord's show, hosted by their co-founder. Going deep on AI in logistics: what it takes to get a model out of a notebook and onto a dock, why the old boring methods still win more often than the shiny new thing, and where quality assurance in warehousing is heading.

deep learning leaders

Posts

05 / off the clock

The Comedowns

Chicago punk / midwest emo. Somewhere between loud and sincere.

  1. Pavement
  2. 916
  3. Take the Wheel
  4. 360 No Scope
  5. Statistically
  6. Way You Aren't

The rest of this site is the day job. This is the night one. The other off-hours constant: four years as a Big with Big Brothers Big Sisters.

06 / contact

Got something worth breaking?

Forward deployed and applied AI roles, technical leadership, or a deployment that will not stay up. Chicago-based, happy to be on site with customers. I read all of it and I answer.