Every company can buy the same models now. The edge moved to where, how, and why they get deployed. I embed with your team, find the workflow where the money actually leaks, build on top of what you already have, and stay until a number you care about changes.
Start a conversationA Forward Deployed Engineer is a customer-facing software engineer who works inside your business rather than at a vendor's office. The role is consultant, product manager, and engineer in one person: discover the real bottleneck, decide where AI belongs in the workflow and where it does not, build on the systems you already run, and own the result until a revenue, risk, or cost number moves.
Palantir invented the role. Anthropic, OpenAI, Databricks, Cursor, and Sierra hire for it now. I do it as an independent, and I explain it in public on the Forward Deploy channel.
Read the full guide: what a Forward Deployed Engineer is and how the method works ↗
Buyer's guide: how to hire a Forward Deployed Engineer, with scope, cost, and timeline ↗
Why the role exists
MIT put the failure rate of enterprise AI pilots near 95 percent. The teams I meet have already tried a chatbot, a copilot, or a vendor demo that looked great in the sales call and went nowhere in the building. The gap is not intelligence. It is deployment.
One loop, run in order. Each stage gates the next. This is the part of the job that is judgment, not code.
The documented process is fiction. I sit next to the person doing the work, pull ten to twenty real completed cases, and walk the request back until I hit the actual bottleneck.
Every step gets one of three answers. A fixed rule becomes deterministic software. A messy judgment call goes to the model. Anything expensive to get wrong stays with a human.
The smallest scope that proves value, on top of the ERP, CRM, or spreadsheet you already run. No migration. Your team's steps stay visible so trust survives.
An eval suite built from your own history. The pass rate is not the deliverable. The failure analysis is. Each failure becomes a fix, a guardrail, or a routing rule to a person.
Shadow mode, then suggestions, then gated actions, then autonomy. Private inference, access control, and audit logs are in the build, not bolted on later.
Businesses measure three things: revenue, risk, cost. We define the number at scoping time and report the before and after in that language and no other.
One proven workflow exposes the next bottleneck up or down the line. The loop runs again. That is how a single deployment compounds across a business.
Three shapes. Scope is directional and the success number is agreed before anything is built.
Two weeks
I shadow the work, map the real process, classify your actual cases, and hand back a triage table with the rough math that says what a deployment is worth. You leave with a funded decision, not a deck.
Six to ten weeks
The smallest build that moves the most work. Shipped inside your systems, with an eval suite built from your history and a rollout that earns trust one autonomy step at a time.
Ongoing
Production is cultural, not just technical. I stay on: monitoring, cost tracking, model changes, and the next workflow the first one exposed. There is no handoff to nobody.
Best fit: a technical product meeting a non-technical team, real volume in the workflow, and an executive who owns the number. If the ask is “we're understaffed,” I will say so on the first call.
Three systems I built and run. Each one replaced a process that lived in spreadsheets, inboxes, and one person's head. Two of them run inside organizations I am part of, so the numbers are mine to stand behind.
The problem
High-touch training programs for Black Girls Code, Upskill Bahamas, Catalyst, and After The Game were running on a spreadsheet, a Google Form, a Zoom account, and a part-time coordinator. Funders asked for outcome data nobody could produce.
What I built
One multi-tenant platform for enrollment, live sessions, attendance, certificates, and reporting. Zoom recordings import themselves. An Eventbrite ticket becomes an account and a magic link. New organizations are database rows, not deploys. An opt-in AI tutor runs per program, with row-level security, a PII audit log, and a weekly automated security check posted to Slack.
What moved
66/66
Every finished cohort to date completed at 100 percent. Five organizations live on one codebase, built in about nine weeks. It now also hosts the Forward Deploy training program for staff.
The problem
Every design and video request was hand-built, and every review round meant three Slack messages per job. Reviewers were drowning and requesters had no idea where anything stood.
What I built
A request row in Notion comes out the other end as finished, delivered assets, with one human approval in the middle. A webhook crawler on Vercel catches the row and asks for missing fields in Slack, a Claude routine plans the work, a Cloud Run studio renders it, and delivery lands in Drive, Notion, Canva, and Slack. Nobody opens a terminal. The Forward Deploy channel publishes from the same system every morning.
What moved
136
Rendered videos across 35 registered projects, plus every still that came with them. One short published daily and a weekly analytics digest, unattended.
The problem
Freelancers and consultants with five clients were paying for tools designed for five hundred employees, and spending more time maintaining the tracker than doing client work. Their Shopify stores also needed fresh ad creative every week with nobody to make it.
What I built
PulsePro.work: projects, tasks, clients, billing through Polar, daily reminders over Telegram, and Claude-generated insights on the dashboard. Beside it, the Ad Engine: it reads a store's Shopify and Meta Ads data, has Claude write five Instagram video ad scripts a week, renders them with HeyGen, and hands back a performance report.
What moved
5 ads / week
End to end in under thirty minutes of machine time, at fifty to one hundred fifty dollars a month in marginal cost per client. The creative problem disappears from the owner's week.
The method, explained in public. A YouTube channel on how Forward Deployed Engineering actually works: discovery mistakes, the triage judgment, evals, the trust curve after launch. Long-form breakdowns plus a 40-second short every day.
Subscribe on YouTube ↗ The FDE Roadmap playlist ↗ Free 30-day roadmap ↗
Start here · 2 min
What a Forward Deployed Engineer does for your business
Play ↗
| No. | Title | Format | Length | Open |
|---|---|---|---|---|
| 02 | The Sommelier's Job: What a Forward Deployed Engineer Actually Is | Long-form | 4:57 | ↗ |
| 03 | The $10M Autopsy: Why 95% of AI Pilots Fail | Long-form | 8:11 | ↗ |
| 04 | 5 Steps to Become a Forward Deployed Engineer | Explainer | 0:58 | ↗ |
| 05 | Why 95% of AI Rollouts Flop: The Sommelier Rule | Short | 0:40 | ↗ |
| 06 | Anthropic Promised Thousands of AI Engineers. Only 86 Exist. | Short | 0:40 | ↗ |
| 07 | Your AI Agent Is Worthless. Here's the 1,000 Ways It Breaks. | Short | 0:40 | ↗ |
| 08 | 5 Job Titles That Are Secretly the Same $280K AI Job | Short | 0:40 | ↗ |
| 09 | Stop Being “Model Agnostic” | Short | 0:40 | ↗ |
| 10 | Every AI Demo Works. 95% of Deployments Fail. | Short | 0:40 | ↗ |
A Forward Deployed Engineer has to ship. This list is pulled live from GitHub: AI tools, ad-ops automation, marketplaces, learning platforms, and voice agents taken from idea to production.
Loading from GitHub…
For nearly five years I was Lead Product Designer for Global Conversion at Netflix: the team responsible for how hundreds of millions of people sign up, pick a plan, and pay across 190 countries. I architected the conversion systems behind the password-sharing crackdown, which turned tens of millions of sharers into paying subscribers and drove billions in new revenue.
Before Netflix I led Growth Design at Coursera. I have founded three companies, invest as an LP, hold an MBA from Quantic, and studied at Stanford and Morehouse. Today I do the job the market is finally naming: embed with a business, decide where intelligence belongs in its workflows, build it, own the result.
Have a workflow that leaks money? Describe the process, who does it, and roughly how often. I will tell you within a day whether a forward deployment fits and what the first two weeks would look like.