[ Peter Brown · Engineer ]
I'm a right-brain engineer with 6+ years of engineering experience, keeping production systems up and building systems of my own. My AI triage service runs in production at work. It investigates. Humans act.
I draw agent systems as a factory floor. The arm is an agent. The conveyor is its software task queue. The one in the hard hat is me. I build software, not robots.
23:00 · Work starts
I wanted less repeating myself and fewer scattered updates. My personal stack has an 11 PM run and a 7 AM brief. Configuring a job is one thing. Knowing where its output went is another.
Personal agent operations platform · overnight run and morning brief (opens in a new tab)
01:12 · The split
After the check, low-risk code, docs and internal tasks go green on their own. Email and spend wait for me at amber. My spend auditor drafts fixes on synthetic data. It cannot apply them.
Ops Margin Agent · drafts fixes only · approval cards use Stripe test mode
03:40 · Recovery
My agent gateway needed its own babysitter. The watchdog tries restart, doctor --fix, then launchd reset, with a timeout on each. It sends a Telegram alert when healed and pages loudly if every step fails.
OpenClaw Watchdog · 5-minute health checks and a 10-case test suite (opens in a new tab)
05:30 · Evidence check
My triage agent claimed a tool had done something it had not. A faithfulness check caught it before it reached anyone. Fluent is not the same as true. I spend my time tuning the check.
AI incident-triage service · labeled cases and faithfulness checks
07:00 · Earned trust
Dawn shows the goal: scoped email earns the auto lane, and spend still comes to me. I want access scoped to the task, limited in time and audited. Trust has to be earned.
LinkedIn · Sep 23, 2026 · scoped, time-limited, audited access (opens in a new tab)
The goal is less babysitting, not less accountability.
I'm a right-brain engineer who also cares whether the demo looks any good. A folder of generated files is not a video. A passing test is not proof of good judgment. I like making things tangible, then checking whether they do what I needed.
Year Up got me through the door at American Express as an intern. I built Slackbots for server restarts and status checks, then became an SRE. That is where I grew up as an engineer: hard problems, good mentors, and real ownership.
Off the clock, I build agents, developer tools, and video workflows. Growing up watching Iron Man probably explains some of that. Repeating myself to an assistant explains the rest. Next, I want to build systems that understand the work well enough to take more of it off my hands.
A missing feature. A crashed gateway. A demo that wasn't a demo. These are the things I built or fixed instead of continuing to complain about them.
Production AI · work
I wanted evidence-backed investigation, not another confident summary. I built an on-call triage service that gathers context and posts findings engineers can review.
Python · Structured outputs · Evals · OpenTelemetry
Open-source contribution
I rely on Hermes every day. When Telegram recovery mishandled auth errors, I reported it and wrote a fix. My fix landed on main. That felt pretty damn good.
Python · Telegram · Regression tests
Open source · recovery
An unavailable assistant is not much of an assistant. I built a macOS watchdog to check the gateway, try bounded repairs, and tell me whether it recovered.
Bash · macOS launchd · Telegram Bot API
Open source · creative
I wanted a demo, not a folder of generated files. Built for the Hermes Creative Hackathon, it turns a GitHub URL into a captioned launch video.
Python · Remotion · FFmpeg · Hermes Agent
Open source · agent skills
I had plenty of coding-agent skills and no coherent workflow. I built a bounded loop for planning, sign-off, implementation, review, repair, and recording the lesson.
Agent skills · Claude Code · Codex · Hermes
Open-source contribution
I wanted xAI billing beside the other providers I used. I added it to CodexBar, then checked that my work had actually landed before saying so.
Swift · Tests
Some work stays in private repos. I can describe the design without publishing the data behind it.
Not everything became a permanent service. These were useful experiments. Two are now archived.
I started by automating server chores. Now I build agents and ask the same questions: what can fail, who notices, and what happens next?
Before July 2019
I came through Year Up, an IT career program. It led to an engineering internship and the chance to learn by doing real work.
July 2019 to Jan 2020
I built Slackbots that replaced manual server restarts and status checks. Small tools, concrete problems. A good place to start.
Since January 2020
I grew through on-call, incident response, observability, runbooks, and self-healing automation. Keeping customer-facing systems up taught me to care about what happens after something ships.
Now
My day job includes an AI triage service in production. At home, I build agents, creative tools, and upstream fixes. I want to carry the reliability judgment forward, not stay inside one title.
| Reliability | Incident response · On-call · SLO and alert design · Postmortems · Runbooks |
|---|---|
| Observability | Dynatrace · Splunk · OpenTelemetry · Grafana · Tracing |
| Cloud & infra | GCP · AWS · Kubernetes · Terraform · Ansible · Linux |
| Code | Python · Bash · TypeScript · FastAPI |
| AI systems | Pydantic AI · Vertex AI · Gemini · Claude · MCP · Evals |
| Agent tooling | Claude Code · Codex · Hermes Agent · OpenClaw |
I write about what agents get wrong, what makes them useful, and why the software around a model matters as much as the model.
Recent LinkedIn posts
Longer reads
Original thesis · Jan 21, 2026Build Palantir Inside Your CompanyI wrote this in January 2026. By May, I was seeing forward-deployed engineer roles across labs, banks and clouds. My argument: build around the actual work.Read on LinkedIn (opens in a new tab)I want to build useful systems close to the people who use them. If your team needs someone who can build the thing and think through what happens when it breaks, I'd like to talk.
Open to opportunities
[email protected]