FDE — Forward Deployed Engineering (The $1M/Year AI Job)

Based on the episode "FDE: The $1M/Year AI Job Explained" from The Startup Ideas Podcast. Host: Greg Isenberg (CEO, Late Checkout) · Guest: Vas from Varick Agents (ex-Meta engineer) Video: https://www.youtube.com/watch?v=zXysLUTLjw4 · Podcast: Apple Podcasts Duration: 51:34 · Published: 2026-07-20

Overview

Forward Deployed Engineering (FDE) is described by Greg Isenberg and Vas as the hottest — and most lucrative — role in AI: a hybrid of deep software engineering and consulting-grade communication that can pay from $150K base plus equity up to $1M/year. The episode's core thesis: frontier intelligence is now a commodity that every company can buy, so the competitive edge has moved from having the AI to deploying it — where, how, and why a company uses it.

Vas traces the role back to Palantir, explains the judgment that decides where AI belongs in a business, details the Audit → Evals → Deployment loop, and lays out a complete 30-day plan to go from zero to production-grade FDE — build, harden, measure, and defend a real agent before ever holding the title.

[!TIP] The companion blueprint referenced in the episode: The FDE Blueprint — https://startup-ideas-pod.link/fde-starter

Chapters

Time	Topic
00:00	Intro
02:03	What is an FDE
04:09	How Palantir Popularized FDEs
06:16	Deciding Where Intelligence Belongs
11:26	What FDEs Earn
14:59	Two Kinds of Judgment: Communication and Engineering
17:38	How the Work Really Gets Done
20:40	Audit, Evaluation, Deployment
22:56	Which LLM to Choose
27:36	Audit: Finding the Workflow Worth Rebuilding
31:47	Evals: Turn Non-Determinism into Evidence
32:57	Deployment: Build on Existing Systems
38:59	The 30-Day Plan Begins
49:13	Final Thoughts

1. The Core Premise: Intelligence Is Commoditized

A new frontier model ships nearly every day; every company can now buy intelligence on tap.
Enterprise clients all use the same stack (Claude Code, Codex, Cursor, GitHub Copilot) — access to capability is no longer a differentiator.
Therefore the edge is deployment: applying general intelligence to each company's specific context — its structure, processes, and exception handling.
The FDE is the bridge between a company's internal reality and the intelligence stack it has access to.

[!NOTE] The episode cites the MIT stat that 95% of generative AI pilots fail — largely attributed to "token maxing" (blanket LLM automation without judgment), which the guest calls a still-fading period in the industry (including a C-suite anecdote of a $10M AI budget burned in 3 months).

2. What Is an FDE

A Forward Deployed Engineer's involvement at a company has three stages:

Understanding the business reality — how work actually happens today (not how it's documented).
FDE judgment — deciding where intelligence belongs and where it does not.
Building the software — deploying the agent/system into production and owning it.

2.1 Origins: How Palantir Popularized FDEs

Palantir built an ontology stack with connectors to software and data lakes, unifying enterprise data in one interface.
Palantir FDEs are deployed on-site (enterprise, military, government), learn the client's workflows, then spin up workflows, dashboards, and agents that solve the client's problems.
The product's beauty was its customizability, not its tech-forwardness — the FDE customized it per client, beating generalized services.
Thesis: if it worked for Palantir in the data age, the AI age will demand it 100× more — every company will need customized agents.

2.2 What the Work Actually Looks Like

The role varies wildly by company:

Palantir-style FDEs mostly spin up workflows/dashboards on the ontology (SQL at most) — light on raw coding.
Other FDEs write full production code on-site with clients, requiring a real software engineering background.
Common denominator: when production breaks, it's on the FDE — you must know who to call, what to do, and how to fix it.

2.3 Compensation

Base from $150K with considerable equity up to $1M/year for the best combinations of consulting + technology.
Described as the hottest role in tech right now — demand outstrips supply.

3. Two Kinds of Judgment: Art + Science

The million-dollar FDE blends two rare skill sets in one person:

Business side ("art"): workflows, cost, incentives, risk, adoption, business value, internal politics — the consulting skill set (McKinsey/BCG/Bain engagement managers are strong here).
Technical side ("science"): models, systems, APIs, data, code, reliability, evals, guardrails, harnesses, post-training, fine-tuning — the software engineering skill set.

[!WARNING] The guest warns against the emerging trend of hiring FDEs who are "neither the best communicators nor the best engineers." The role demands the best of both, not an average of both — "if you understand art and you understand science and you can speak both, you have what it takes to become the million-dollar FDE."

4. How the Work Really Gets Done

The documented process is very rarely the real process. Example: an "email arrives" workflow — 40+ senders, no consistent formatting (PDFs, screenshots, Excel, forwarded threads), half exceptions, and the real routing logic lives in one person's head.
The FDE sits with the people doing the job (often multiple people), observes, interviews, and coaxes out undocumented exceptions — this is where the bulk of the time goes.
On-site presence matters: a 1-hour scheduled walkthrough gives you the perceived job; a full day on-site shows you the real job, including exceptions that never make it into an SOP (the McKinsey "sit with the miners" method).

5. The FDE Loop: Audit → Evals → Deployment

The three-phase operating loop; each stage is a prerequisite for the next, and every completed loop makes the next workflow clearer ("you're not 10x-ing one workflow — you're 100x-ing the business").

5.1 Audit — Find the Workflow Worth Rebuilding

Collect context, trace findings, identify bottlenecks, repetitive work, and judgment points → produce an operating map.
Prioritize high-volume workflows where the improvement is large enough to matter.
Deliver a priority map / ROI matrix ("what's worth automating vs not") — often valued at 10× the fee paid ("better than McKinsey").
Selling the audit: charge for it, or do it for free to de-risk first customers — "your first few customers are worth more to you than you are to them."
Tip from LCA (Greg's agency): rebrand "audit" as a "design sprint" — clients have an allergic reaction to the word.

5.2 Evals — Turn Non-Determinism into Evidence

Goal: prove the system behaves correctly with the right data, required steps, expert alignment, and safety to act.
Deterministic tasks: easy (e.g., categorize 10K historical emails → eval set).
Non-deterministic tasks (e.g., presentations): build a golden dataset from as much prior output as possible (5K previous presentations), encode what matters (logo placement, styling), and accept you'll never reach perfect — always bake in human-in-the-loop feedback to keep improving the harness.
Produce an evaluation report: "50 runs, 41 passed; of the 9 failures — 5 missing data, 4 wrong record pulled" → use it to improve the system.

5.3 Deployment — Build on Existing Systems

Integrate with what already exists — don't force migrations. Example: a client spent years and millions moving to NetSuite; an AI solution demanding another migration gets "get lost." Build on top of NetSuite and integrate it with Salesforce, SAP, Concur, Expensify, Gong, Workday instead.
Graduated rollout: controlled environment → shadow mode → increasing autonomy → production.
De-risk for the humans: people at companies don't want to get fired — they want to get promoted. Frame value as driving value cost-effectively (revenue, reach, efficiency) so the executive sponsor gets credit at review time.
FDEs are viewed as risk; mitigate by proving value early (free audit) and only getting paid after measurable value.

6. Which LLM to Choose

Model agnosticism is a company-level strategy (Varick Agents deliberately stays swappable to keep accuracy up and costs down — "don't monopolize your inference layer").
For an individual FDE starting out: do the opposite. Pick one model + one agent-building platform (OpenAI, Claude/Agent SDK, etc.) and get very good at it — that's the foundation. Expand later (Kimi 3, GLM, open-source models, proprietary harnesses).
The sommelier analogy: don't hand everyone a Pinot Noir — understand the customer's palate first. "Your ability to determine what model is best for a task relies on your understanding of various models... but don't put the cart ahead of the horse."

7. The 30-Day Plan: Doing the Job Before You Have the Title

Condensed from a year of real FDE work (Vas learned it as an ex-Meta engineer with no consulting background). Spaced as needed — "30 days" is a framework, not a deadline.

Week 1 — Build an Agent That Completes a Real Loop

Pick one real enterprise back-office workflow (finance, HR, procurement, logistics, sales) at maximum granularity; build an agent for it.
Agent definition: "if I give you a task, can you solve it in as much detail and with high enough accuracy as possible — without relying on someone prompting perfectly."
One day each: agent looping, tool usage, guardrails, context & memory, audit trail.
Audit trail is non-negotiable: if you can't show the client what the agent did, they will never trust it — log every trace (a software engineering problem that sets you apart).
Checkpoint: a working agent with tools, guardrails, deliberate memory, and a full audit trail for one task.

Week 2 — Turn the Demo Into a System That Can Recover

Engineering hardening: defined JSON schema (not freeform text), schema validation, failure modes / exception handling.
Build for the unhappy paths: "there's only one way something can go right, but a thousand ways it can go wrong. If you only build for the way it goes right, you're worth nothing."
Deep client understanding (how things actually fail) makes exception handling far more effective.

Week 3 — Make It Measurable and Economically Viable

Retry logic, golden dataset for evals, continuous improvement.
Try cheaper / less frontier models for subtasks (e.g., Gemini Flash-class models) to cut cost.
Measure across the three buckets that matter to every business: revenue uplift, risk mitigation, cost savings.
Checkpoint: an evaluated agent with known failure modes, measured costs, and a golden dataset.

Week 4 — Defend the System Like an FDE

Build the business case: pain points, why AI belongs, architecture, iteration history ("accuracy went from 70% to 95%"), economics.
Rehearse it twice: as an engineer (architecture, decisions) and as a VP (problem, outcome, evidence, risk).
Pitch the agent to real businesses — they'll tell you point-blank if the economics and framing are right.
Day 30: you understand FDE and have evidence you can do it — the pitch that gets you in the door.

Week	Theme	Key Deliverables
1	Build	Working agent with tools, guardrails, memory, audit trail
2	Harden	JSON schema, failure modes, exception handling
3	Measure	Evals, golden dataset, cost optimization, 3-bucket metrics
4	Defend	Architecture + business case, pitched to real businesses

Key Takeaways

Intelligence is now commoditized → the edge lives in deployment — the FDE's job.
The FDE role traces to Palantir: engineers embedded on-site, customizing per client.
Strongest FDEs = deep technical skill plus consulting-grade communication — the rare art + science combo worth up to $1M/year.
The operating loop is Audit → Evals → Deployment; each improved workflow makes the next one clearer.
A year of learning can be condensed into a 30-day plan: build, harden, make measurable, then defend.
Still early: "doing the job before you have the title" is the entry strategy.

Resources

FDE Blueprint (30-day guide): https://startup-ideas-pod.link/fde-starter
Varick Agents: https://www.varickagents.com
AI Forward Deployed Engineers (learn.varickagents.com/fde-in-30-days): https://learn.varickagents.com/fde-in-30-days
Vas on X: https://x.com/vasuman
Greg Isenberg on X: https://twitter.com/gregisenberg
IdeaBrowser (startup ideas/trends): https://www.ideabrowser.com
Late Checkout / LCA: https://latecheckout.agency