Agentic engineering, done right
I help engineering teams adopt agentic engineering the right way, with the rigor that makes machine-generated code safe to ship. The opposite of vibe coding.
Reject expired invite links
Issue 412, in progress
What I do
I coach engineering teams through the shift from writing code to specifying intent. Claude Code & Cursor rollouts, spec-first workflows, and supervisory protocols — so your seniors stop drowning in machine-generated code.
Autonomous pipelines built on embry0, my open-source engine, and wired into your dev workflow. An issue becomes a reviewed pull request: triage, developer, reviewer and QA agents in sandboxed containers, with a person approving what matters.
See embry0I turn fuzzy PRDs into AI-ready specs: state machines, decision tables, hallucination-proof test suites. The same standards safety-critical avionics teams apply to flight hardware — applied to AI-generated code.
Hands-on architecture and development. When you need a senior who can ship the thing, not just talk about how to ship it. Backend systems, agent infrastructure, secure networking, full-stack.
Packages
Instead of months of readiness workshops, you see the fix running on your own data before you commit to anything.
Agents your staff message in Microsoft Teams, each limited to what that person is allowed to do.
What you getAn issue goes in, a reviewed pull request comes out, with evidence that the change works.
What you getThe work between a request and a signed deal, done by agents your reps supervise.
What you getThe same work for less: every request sent to the cheapest model that can do it well.
What you getA system built around one problem the other packages don't cover, on the stack you already run.
What you getOpen source, MIT licensed
A self-hosted engine I built and now publish in the open. A triage agent plans the job, a developer agent writes the change, a reviewer sends it back until it is right, and a QA agent boots your app to check it in a real browser. You run it from a dashboard that can look like a pixel-art guild or a plain operations console.
embry0, the guild hall
The guild hall. Every agent has a desk. You see who is busy, what is on offer and what is waiting for your review.
An issue
Orchestrator
Sandbox, created for this job and destroyed after it
non-root, capabilities dropped
Developer and reviewer loop until the reviewer approves. Your tests then run on the old code and the new, and only new failures stop the job. QA runs when the change touches an app it can boot.
Credential proxies
A reviewed pull request
You, at any step. Agents stop and ask when they are unsure. Answer from the dashboard, Telegram or a GitHub comment.
Any model, per agent. Claude, Grok, OpenAI-compatible providers, or a local model on your own hardware.
Results
A freight brokerage. I designed and led its AI backend, and built the pricing models, the email pipeline and the deploy platform.
12.5% → 8.1%
Typical error when predicting what a dry-van load will cost to cover, against the industry rate benchmark.
Backtest on 8,414 loads over 90 days. The window used to choose the design was kept separate from the one used to score it.
23.8% → 9.0%
Typical error on power-only loads, against the per-mile rule.
1,317 loads over a year, with the model refitted every month and scored only on loads it had not seen.
3× faster
Email classification after constraining the model to a structured reply: 756 ms instead of 2,254 ms, agreeing with the stored labels 96.6% of the time.
Holdouts of up to 5,000 threads. It caught 98.2% of new quote requests.
About 6 million
Email classifications into nine business categories by an open-weight vision model on the company's own GPUs.
No email left the building. Attachments are read as images, and text is redacted before it is used for training.
432 production releases in about six months, each one through a canary and a verify window with automatic rollback. The client is not named. Every figure comes from a written backtest or evaluation report.
How I think about this
AI hasn't made engineering easier. It has moved the bottleneck upstream. Code is disposable; specifications are the product. And the supervisor who doesn't understand what the system is doing is the one who crashes the plane.
Good fit
If that sounds like you, we should talk.
Engineering teams of 5–30 at companies that ship software as a core part of the business.
Teams already using or seriously evaluating Claude Code, Cursor, or similar AI coding tools.
Senior engineers who feel like they're drowning in machine-generated code.
About
I spent my career in domains where software failure isn't an inconvenience: avionics communication platforms, embedded firmware for industrial control systems, and secure network infrastructure for high-assurance environments.
The through-line: every system I've shipped has been one where "kind of working" is unacceptable. Safety-critical hardware. Regulated environments. Networks carrying traffic nobody can afford to leak. AI-generated code in production needs the same standard.
Eastwood AI is how I help engineering teams hold that line while moving faster than ever — through coaching, custom agent infrastructure, or hands-on engineering work.
Free 30-minute intro call. No pitch deck, no sales pressure. We figure out together whether we're a fit.