Forward Deployed Engineer
A forward deployed engineer who gets AI into production.
I embed inside your business to identify high-value opportunities, build production AI systems, and leave your team able to run them without me.
This is the pipeline behind the contact form on this site. The dashed edge is the point: low-confidence messages route to a person, and the agent has never been allowed to send.
Shipped in production with
- TypeScript
- Swift
- Next.js
- Vapor
- PostgreSQL
- Fly.io
- Vercel
- Playwright
The problem
AI is everywhere. Production outcomes are not.
Most companies have access to powerful AI tools. Far fewer have turned them into measurable business impact.
~95%
of enterprise AI pilots showed no measurable P&L impact, per The GenAI Divide (MIT NANDA, 2025): 52 interviews, 153 survey responses, 300 public deployments.
Its critics are right that it counts only P&L, so efficiency and cost gains fall outside it. Treat it as directional, not precise. I would rather hand you the caveat than the headline.
The challenge is not choosing the right model. It is knowing where AI creates value, connecting it to your data and workflows, and shipping something people actually use.
Pilots fail for three reasons, and none of them are the model.
Nobody mapped how the work gets done
The system was built for the tidy version in the process document, which is never what happens on a Tuesday.
Nobody could measure whether it was right
So “the demo looked good” became the standard for going live, and nobody could tell whether it was improving.
It was trusted on optimism
One visible mistake in front of a customer killed the project, plus the appetite to try another.
What is a forward deployed engineer?
A senior builder embedded with your team.
I work across business, product, data and engineering to turn AI opportunities into working software. Not slide decks, not disconnected prototypes, and not staff augmentation.
The distinguishing skill is the second half. Plenty of people can run the workshop. Fewer can ship what comes out of it, prove it works, and hand it over so it survives without them.
I help you move from AI ambition to AI in production.
Shipped
Not a logo wall. Things that are running.
I don't have enterprise client logos to show you. I have systems in production with people depending on them, and the numbers behind each one.
How this works
Start with an audit. Scale as value grows.
I meet you where you are, with the smallest engagement that could tell you something true as the starting point.
Start here
2 weeksThe Audit
Find out whether this is worth building at all.
Best when you have a process that is obviously expensive but nobody has established what it costs or whether AI should touch it. I map the real workflow and put numbers on it.
Use this when you need
- A defensible answer on whether to proceed
- A map of how the work really happens, exceptions included
- What it costs you today: hours, money, error rate
- A written plan for proving it works before you trust it
Scale up
4–8 weeksThe Build
Take one process from nothing to running on real work.
Best when the opportunity is already clear and what you need is the system itself, built to survive production rather than to demo well.
Use this when you need
- A working system handling the process end to end
- The unglamorous parts: bad input, outages, duplicates, confident wrong answers
- A test set from your real cases, and a published score against it
- Alerts, a shut-off, and runbooks so your team operates it without me
Full journey
OngoingEmbedded
I work as part of your team and stay on the hook.
Best when AI is a standing priority rather than one project, and each workflow should land faster than the last because the foundations are already there.
Use this when you need
- Everything in The Build, across a backlog of processes
- Shared foundations, so each build compounds
- Ongoing testing as production throws up cases nobody predicted
- Reviews of what has earned more autonomy, and what has quietly lost it
Standardising on Claude?
The agent on this site is built on it.
Not a partnership badge — a running system you can inspect. Triage, fit scoring and reply drafting on anthropic/claude-sonnet-5, routed through the Vercel AI Gateway so the provider stays a configuration value rather than an architectural commitment. The eval harness is provider-agnostic, which makes switching a measurement instead of a rewrite.
The operating model
How the work creates value.
Three steps, in this order, every time. The order is the method — most projects skip straight to building, which is exactly why most projects fail.
Find the value
Two weeks of observation before anything is built.
I sit with the people doing the job and follow it through the systems it actually lives in — the two senders who get special handling, the step that exists because software from 2011 can't take a decimal point, the person who knows all of it and is on holiday. You get a map of the real process and hard numbers on what it costs you today.
Sometimes this ends with me telling you it isn't worth automating. That costs you two weeks instead of two quarters.
Ship the value
An agreed definition of correct, then a system scored against it.
We write down what “correct” means for this task and collect real cases from your business where the right answer is already known. The question stops being “does this seem good?” and becomes “it got 94 of these 100 right, and here are the 6 it missed and why.”
This is the step almost everyone skips, and skipping it is the single biggest reason their projects died.
Multiply the value
It runs beside your team, then your team runs it.
Full capability, full workload, nothing at risk — recording what it would have done so you can compare against your own people on the same live traffic. Responsibility is granted one rung at a time, and each rung is bought with evidence from the one below.
You keep the documentation, the tests, and the ability to run it without me. Including the ability to fire me.
What I build
Production systems tied to real workflows.
Not demos, not proofs of concept that live in a notebook. Common outcomes include:
Agents that do a real job
Scoped to one workflow, with a named owner and a shut-off.
Workflow automation
The intake, routing and follow-up a person currently does by hand.
Internal tools
The thing your team keeps rebuilding in a spreadsheet.
Retrieval over your own documents
Answers grounded in your material, with citations back to it.
Evaluation harnesses
A golden dataset and rubrics, so improvement is measurable.
Data plumbing
The unglamorous integration work that everything else depends on.
Permission and audit models
What the system may touch, and a record of everything it did.
Native mobile
When the workflow lives in someone's hand, not on a desk.
Proof, not badges
The system on this site publishes its score, including what it fails.
This is the contact form's triage agent, measured against a golden dataset I wrote. The numbers below are read from the eval report at build time, so they cannot drift away from the truth.
- 12/14
- cases passed, scored on 6 independent dimensions at once
- 2
- failures, published with the cause named rather than quietly removed
- 10.1s
- median latency end to end, measured not estimated
- $0.015
- per case, so the running cost is a known number
It classifies intent, scores fit, and drafts a reply on anthropic/claude-sonnet-5. It has never been allowed to send. Below a 0.6 confidence gate its answer is recorded but not acted on, and the message comes to me instead. A vendor under pressure adjusts the test until the score looks better — assume that is what happened whenever someone shows you a perfect one.
Why me
Built to put AI in production.
One engineer across the whole stack. Nothing is handed across a boundary where the context goes to die, and the person who mapped your workflow is the person who ships it.
Production is all I do
Full-stack across web and iOS, plus the backend and the hardware when the problem needs it. I've shipped production TypeScript and production Swift, so I can follow a workflow into whatever system it actually lives in and keep going past the API boundary.
I finish things
FalconryLab is a real iOS product on the App Store, built by embedding in a niche domain and learning how practitioners actually work — the same motion this role runs, self-directed. The agent on this site is the other half: shipped, measured, still running.
I can explain it
The differentiator in this job is defending the system to an engineer and to a non-technical executive in the same afternoon. I've been writing for as long as I've been building, and this whole site is the sample.
FAQ
Common questions
Put an engineer where the work happens.
Describe the process that is costing you the most. If I don't think it's worth automating, I'll tell you that instead of selling you something.