Varick — Forward Deployed Engineering
Varick
Varick’s guide to

AI Forward Deployed
Engineering

The blueprint to breaking into FDE in 30 days.

SCROLL ↓

Every company can now buy intelligence.

The same foundational capability is becoming available to anyone able to pay for it.

If everyone can access it, intelligence alone cannot be the moat.

So where does the advantage go?

Into deployment.

The edge is no longer who has intelligence. It is where, how, and why they use it.

Someone has to decide where intelligence belongs.

That person is the forward deployed engineer.

Business reality

How the work actually happens today — tools, people, exceptions.

FDE judgment

Where intelligence belongs, and where it does not.

Deployed AI system

Working software carrying real responsibility inside the business.

FDEs are in demand because they control how intelligence enters the business.

The role requires two kinds of judgment that rarely exist in one person.

The commercial worldThe technical world
workflowsmodels
costsystems
incentivesAPIs
riskdata
adoptioncode
business valuereliability
FDE

The FDE turns business understanding into working software.

First, understand how the work is really being done.

The documented process is rarely the real process.

the FDE traces every step

Click any step — the “simple” version hides the real work.

Then decide how the work should operate when intelligence is built into it.

an exception escalates to a person Deterministic software intake · validation · routing The agent acts gathers · decides · drafts A human approves one clear decision Record updated the ERP, without re-keying EVIDENCE LOG — EVERY STEP RECORDED

An FDE understands the current workflow, designs the AI-native workflow, and builds the system that connects the two.

ONE RUN · END TO END
invoice_0417.pdf
DROP TO INGEST

Finally, the agent can be deployed over the existing systems.

Intake validated
Agent drafts
Human approves
Record posted
RUN LOG
document received · invoice_0417.pdf
intake · fields parsed, no duplicates found
agent · pulling the matching PO from the ERP
agent · cross-checking contract terms
agent · drafting the record
paused — waiting for human approval
approved · by you, just now
posting to the ERP — no re-keying
record posted · evidence log complete

✓ Record posted — every step of the run captured in the evidence log.

The job has three parts.

Audit

identifies the right problem and maps reality.

Evals

prove the system behaves correctly.

Deployment

makes it work inside the business.

Each stage earns the right to the next.

Audit: Find the workflow worth rebuilding.

The audit determines what should be automated before anyone begins building.

1Collect context
interviews emails spreadsheets SOPs systems approvals exception paths
2Trace FDE findings
bottlenecks repetitive work judgment points required systems failure modes automation worth doing
3Produce the operating map
THE AUDIT ARTIFACT — ONE OPERATING MAP
Current-state workflowhow the work actually happens today
Future-state workflowthe same work, rebuilt around AI
Selected use caseone workflow, chosen for value
Boundarieswhat the system may and may not do
Expected business valuehours, cost, and errors — quantified

Decide what should be automated — and what should not.

a step in the audited workflow DETERMINISTIC SOFTWARE When the rules and inputs are predictable. AN AGENT When the objective is clear but the inputs, path, or required actions vary. A HUMAN IN CONTROL When the decision carries material ambiguity, accountability, or irreversible consequences.

Prioritize lengthy, high-volume workflows where the improvement is large enough to matter.

Evals: Turn non-determinism into evidence.

THE EVALUATION CASES
Right data
Required steps
Matches expert
Safe to act
The normal case
The edge case
Incomplete information
→ human
An ambiguous request
→ human
A high-risk action
→ human
✓ expected behavior  ·  ✗ failure, with the evidence kept  ·  → human  escalated by rule
THE EVALUATION REPORT
Pass rate
41/ 50graded runs passed — 82%
Failure categories — 9 runs
missing data5
wrong record pulled4
Operating rules
CONFIDENCEbelow 0.80, the system must escalate
ESCALATIONhigh-risk actions always reach a person
READINESSpilot — with human review on every action

Deployment: Make it work inside the business.

existing data internal APIs identity & permissions SANDBOX human review logs alerts rollback production users THE SYSTEM PROTOTYPE IN PRODUCTION
1Integrate with what already exists

Build over the current data and systems rather than beginning with a large replacement project.

2Test inside a controlled environment

Use a sandbox within the company’s infrastructure to run, inspect, and debug safely.

3Increase autonomy gradually

Begin with the smallest useful action and grant more authority only after the system proves reliable.

Deployment is where software begins carrying operational responsibility.

Audit Build Evals Deploy Observe Improve

An FDE owns the path from messy workflow to trusted production system.

THE 30-DAY PLAN ↓
The 30-day roadmap

How to become an FDE in 30 days.

Most successful FDE candidates come from consulting, product management, or software engineering.

COMPLETE THIS ROADMAP IN PARALLEL WITH APPLYING AND INTERVIEWING

CONSULTANTS AND PMS — CHOOSE TWO PROJECTS AND TAKE THEM DEEP
Production-ready AI agentRAG pipelineEval frameworkMCP integration
WEEK 1 · DAYS 1–7

Week 1: Build an agent that can complete a real loop.

By Day 7, the agent should complete one useful workflow and expose every step it takes.

WEEK 2 · DAYS 8–14

Week 2: Turn the demo into a system that can recover.

By Day 14, the agent should produce predictable outputs and resume after failure.

WEEK 3 · DAYS 15–21

Week 3: Make the system measurable and economically viable.

By Day 21, you should know how the system fails, how much it costs, and whether it is improving.

FINAL WEEK · DAYS 22–30

Final week: Defend the system like an FDE.

By Day 30, you should be able to explain the system to both an engineer and a non-technical executive.

Do the job before you have the title.

On Day 30, you should not merely understand forward deployed engineering. You should have evidence that you can do it.

DAY 7Working agent
DAY 14Recoverable system
DAY 21Evaluated system
DAY 30Complete FDE case study