Forward Deployed Engineering

Everyone can buy the same intelligence.
The advantage is in how you deploy it.

Frontier models are now a commodity — a business in Port Moresby can call on exactly the same intelligence as one in San Francisco. So the edge no longer lives in the model. It lives in the judgment of where that intelligence belongs in a real workflow, and the engineering to put it into production safely.

That work has a name: the forward deployed engineer — someone who sits with a business, learns how it actually runs, and ships AI into the systems it already uses. PNG Cloud is that practice, built for Papua New Guinea.

What a forward deployed engineer actually does

The role was popularised at Palantir: rather than shipping software and hoping it fit, they embedded engineers on-site with each client, learned the real workflows, and tuned the system to that specific business. An FDE is consulting for the software age — half engineer, half translator of business reality.

Embed, don't guess

Sit with the people doing the work and watch how it really happens — not how the org chart says it does. The insight comes from observation, not a requirements document.

Decide where intelligence belongs

The core judgment: which steps genuinely need a model, and which should stay deterministic software or simple rules. Most failed AI projects reach for a model where an if-then would do.

Ship into what exists

Deploy on top of the systems a business already runs on, with full audit trails and a human in the loop — so the people relying on it can trust what it does.

Two kinds of judgment

The rare, valuable combination is a fluent communicator who reads the business and an engineer who ships production systems. Art and science in one person.

95%

The figure Vas cites from MIT: roughly 95% of generative-AI pilots fail to reach production. Not because the models aren't capable — because the intelligence was pointed at the wrong part of the workflow, with no measurement and nothing to build trust. Forward deployment is the discipline that closes that gap: the right step, made measurable, deployed where the work already lives.

The loop: audit → evals → deployment

Forward deployment isn't a one-off build. It runs as a loop, and each workflow it improves makes the next one clearer.

01 · Audit

Find the workflow worth rebuilding

Map how the work is really done and locate the one process where better decisions would move revenue, risk or cost. Not everything deserves an agent — the audit is where that judgment happens.

02 · Evals

Turn non-determinism into evidence

A model's output is fuzzy; a business needs proof. Evals convert "seems to work" into measured pass/fail against real cases, so improvement is visible and regressions are caught.

03 · Deployment

Build on the systems that exist

Put it into production on top of what the business already runs — with audit trails and human approval — so it's trusted, defensible, and actually used.

Doing the job before you hold the title

In the podcast, Vas compresses a year of learning into a 30-day path — build something real, then harden, measure and defend it. It's a useful shape whether you're learning the craft or judging whether a deployment is production-ready.

Week 1 · Build

An agent that completes a real loop

End to end, doing one genuine task from start to finish — not a demo.

Week 2 · Harden

Schemas, failure modes, exceptions

Make it survive the messy inputs and edge cases the real world throws at it.

Week 3 · Measure

Revenue, risk and cost

Instrument it so its value — and its failures — can be seen and counted.

Week 4 · Defend

Like an engineer and a VP

Stand behind it on both the technical detail and the business case.

Forward deployment, for Papua New Guinea

PNG Cloud, powered by HDC, is this idea applied to the Pacific. The same frontier intelligence is available here as anywhere — the gap has always been deployment into PNG's real conditions: kina, PNG GST and SWT, local banking formats, and the realities of Pacific connectivity. So rather than sell generic software, we build products that already carry that judgment inside them.

● Live

Hero

Online accounting engineered around PNG tax law — SWT payroll, GST, kina pricing — served at the edge with the database close to PNG. AI is deployed where it earns its place: reading receipts, drafting from your own figures, never in the ledger math.

Open Hero →
● Live

Toktok

Customer messaging with an AI agent grounded strictly in a business's own help articles — answering in the channel PNG actually uses, and handing a real person the conversation the moment judgment is needed.

See Toktok →
Partner

Your workflow

Have a process where better decisions would move the needle? That's an audit. We start by finding the one workflow worth rebuilding, prove the value, then deploy it into the systems you already run.

Start a conversation →

Where this thinking comes from

The framework on this page — the FDE role, the audit → evals → deployment loop, and the 30-day path — is drawn from a conversation worth watching in full.

“FDE: The $1M/Year AI Job Explained”

Greg Isenberg (The Startup Ideas Podcast) in conversation with Vas of Varick Agents. The ideas here are theirs; the PNG framing is ours. Credit where it's due — and the original is worth your time.

Watch on YouTube ↗ Varick Agents ↗