Dylan Ironside
Résumé PDF

Applied AI engineer · Coralville, Iowa · Remote (US)

I build agent systems, and direct the agents that build them.

I created Super Hornet, a multi-provider AI agent platform, and used it to automate radiology second-opinion drafting at Braid Health, where a radiologist approves every case.

Download résumé PDF · 1 page Email me GitHub

Open to applied AI and agent-tooling roles.

Decision log

ADR-0201Accepted

Context compaction

Was Compact when a token estimate says the context is almost full.

Now Compact only after the provider rejects the actual request as too large.

Estimates disagreed across OpenRouter, Claude, and Codex. The provider’s answer is the only one that can’t be wrong.

ADR-0205Accepted

Agent concurrency

Was Start a run only if its estimated memory reservation fits.

Now Run as many agents as the user’s own setting allows.

A hidden reservation quietly cut the worker to one run at a time. Real resource failures now surface instead of silently queueing.

ADR-0012Accepted

Durable agent runs

Was Hold each model stream open inside the web request.

Now Queue runs in Postgres; workers own them and log every event in order.

A deploy used to kill live streams. Now work survives deploys and closed laptops, and replays in order.

Braid HealthIn use

Radiology second opinions

Was Hours of copy-and-paste into a chatbot for every case.

Now An agent drafts in seconds; a radiologist approves every case.

Prompts stop on mismatched patient records and flag missing reports. Data stays in an isolated workspace under a HIPAA BAA.

Four calls I made, from my architecture decision records and Braid’s clinical workflow.

By the numbers

tokens in one week, as Super Hornet’s only user
10B+Self-reported
architecture decision records I’ve written
336Super Hornet repository
automated tests guarding agent-written code
~8,700Super Hornet repository
commits since the first one on April 30
3,937Super Hornet repository
points on the Show HN front page
265news.ycombinator.com
GitHub stars on Hacker News for macOS
428github.com

As of October 4, 2026. The Super Hornet repository is private; I’m glad to walk through it on a call.

§ 01

Selected work

Creator · Independent project

2024 – Present

  • StatusSole user; closed beta planned
  • Scale~450K lines of TypeScript
  • Models400+ through OpenRouter
super-hornet.com

Super Hornet

A multi-provider AI agent platform: any model gets its own Linux machine, a shared browser, and a team of subagents. It replaced Claude Code, Codex, and Cursor in my own daily work. Every line was written by AI coding agents under my direction.

Providers
OpenRouter (400+ models), the Claude Agent SDK, and Codex behind one model catalog, one tool contract, and one durable event stream, using OpenRouter’s routing, fallbacks, and zero-data-retention filters.
Durability
Server-side workers own model calls and tool loops, so a run survives a closed browser or a deploy and resumes from an ordered Postgres event log.
Compaction
Replaced token-estimate compaction, which behaved differently on every provider, with one reactive policy: when a provider rejects a request as too large, the thread is summarized hierarchically and the run resumes on the same model.
Evals
Given a failing conversation’s ID, an agent with its own test account pulls bounded diagnostics, reproduces the failure in the live app, and proposes a fix. Live suites exercise all three providers.
Security
A microVM per conversation, per-operation GitHub credentials checked server-side, encrypted secrets, SSRF-protected services, and approval before compute tools run.

Senior Technical Support Specialist

Oct 2022 – Present

  • Company~10-person AI teleradiology startup
  • RoleHalf engineering, half support
  • RemoteCoralville, Iowa
braid.health

Braid Health

I built the AI workflow behind Braid’s $199 radiology second opinions, on Super Hornet. Case drafts that took hours of manual chatbot work now take seconds, and operations leadership adopted it after a demo.

How a second opinion moves through the workflow
  1. PatientIntake form and prior reports
  2. AgentDrafts the case presentation, flags missing reports
  3. RadiologistReads the images, writes the second opinion
  4. AgentTurns the read into a patient-friendly report
  5. RadiologistApproves or edits before it reaches the patient
Guardrails
Prompts that flag missing reports, stop on mismatched patient records, and keep report findings separate from patient claims. Data stays in an isolated OpenAI workspace under a HIPAA BAA.
Replication
I operate Braid’s CouchDB replication on GKE and lead outage response: a sync engine per organization (~10) moves X-rays from clinic hubs through a PHI-filtering proxy to radiologists, across tens of thousands of per-user databases. I built the Datadog synthetic tests and alerts, and added a revision cap after runaway documents halted syncing.
Bucket+
A desktop app three support staff use daily to bulk-download, unzip, dedupe, and sort patient imaging from Google Cloud Storage, replacing one-by-one downloads of unlabeled folders for hundreds of patients.
Support
Most of the 20–50 daily Intercom tickets on a team of one or two: bugs tracked in Linear, patient records repaired, new clinical customers onboarded, and the Intercom workflows and macros behind it all.

§ 02

How I work

I don’t hand-write code. AI coding agents write all of it, and I own everything that decides whether it’s any good.

  1. 01

    Decide

    Every significant change starts as a decision record: the context, the options, the call, and what it costs. Superseded records stay, so the reasoning is never lost.

    336 records

  2. 02

    Delegate

    I scope the work, pick the model for it, and run several agents in parallel, each in its own sandbox with only the credentials its task needs.

    3,937 commits in five months

  3. 03

    Verify

    Nothing merges on trust. Quality gates and tests run on every change, live suites exercise every provider, and an agent reproduces reported failures in the running app.

    ~8,700 tests

  4. 04

    Operate

    I stay on the hook after launch: synthetic tests and alerts on Braid’s replication, first response when it breaks, and a fix that keeps it from breaking the same way twice.

    First responder at Braid

§ 03

Experience & toolbox

  1. 2024 – nowCreator & AI Engineer, Super HornetIndependent project
  2. 2022 – nowSenior Technical Support Specialist, Braid HealthAI workflows, internal tools, infrastructure operations, support
  3. 2022Assistant Blockchain Developer, NFTNTWAX NFT backends for Trailer Park Boys and King Kong
  4. 2016 – 2021Support & salesTaxAct, Businessolver, Verizon (Salesforce), Best Buy, Ruffalo Noel Levitz (CRM)
  5. 2020 – 2021Kirkwood Community CollegeAssociate’s degree coursework
  6. 2018Coe CollegeLiberal arts coursework
AI & LLM
Agent harnesses · Multi-provider routing · Context management · Prompt and guardrail design · Directing coding agents
AI tools
OpenRouter · Claude Agent SDK · Vercel AI SDK · OpenAI · Claude Code · Codex · Cursor
Stack
TypeScript · Python · Swift / SwiftUI · Next.js · React · PostgreSQL · CouchDB
Infrastructure
Docker · Kubernetes (GKE) · E2B microVMs · Cloudflare
Quality & ops
Playwright · Vitest · Self-hosted CI · Datadog · Sentry · Linear · Intercom

§ 04

Let’s talk

Email is the fastest way to reach me.