By the end of this lesson you will be able to define the FDE role, describe a typical week, and name the six skill areas this course trains.
By the end of this lesson you will be able to define the FDE role, describe a typical week, and name the six skill areas this course trains.
What problem, for whom, measured how?
Sit with Priya's intake team; pull fax and EHR timestamps. Baseline: 71 hours median.
Monday: review pilot metrics with the ops lead
Discovery forever: timebox it, build something by the end of week 1
Your company just signed Brightline Health, a clinic network where faxed referrals take three days to reach the patient record. Dana Ruiz, VP Operations, wants intake faster by June, not a demo. You're the engineer being sent in.
A Forward Deployed Engineer (FDE) owns the outcome at a customer and writes whatever production code it takes to get there. The customer's problem decides what you build, not a roadmap.
That sentence has two halves, and both matter. "Owns the outcome" means you're judged on whether Brightline's intake actually gets faster, not on whether you closed your tickets. "Writes production code" means you're an engineer, not an advisor: the extraction pipeline, the review queue and the script that pulls the fax archive are yours to build, run and debug in an environment you don't control.
Palantir made the title well known by sending engineers on-site to connect its platforms to messy customer data. Two lessons stuck: deployments generate the best product insight, and a prototype on the customer's data in week two beats a slide deck. With LLMs, someone still has to find the right task, connect Claude to the data, prove it with evals, and ship it.
That last list is why the role has grown with LLM products. A model that is excellent in a demo still needs someone to decide which of Brightline's workflows it should touch, get the faxes out of a system nobody has exported from before, measure whether extraction is accurate enough for patient records, and convince an IT director that patient data is handled properly. None of that is in the model. All of it is FDE work.
The diagram above walks the same loop you'll run all course: discover, build, deploy, measure, feed back. A week rarely follows it in order, but every day touches one step. Here is a typical week once a pilot is running:
| Day | What happens |
|---|---|
| Monday | Review pilot metrics with the ops lead. Two edge cases fail. |
| Tuesday | Pull 30 days of tickets from their API; add the failures to the eval set. |
| Wednesday | Fix prompt and retrieval: eval accuracy 81% → 90%. Ship to the pilot group. |
| Thursday | Write a one-page data-flow doc for their security team. |
| Friday | Tell your product team: "Three customers asked for X. Here's my workaround." |
Notice the mix. Two days are code, one is evaluation, one is a security document and one is product feedback. If you only enjoy the code days, the role will frustrate you; if you enjoy the whole week, it's one of the best jobs in software.
The second. The diagram answers a question nobody has asked yet. The measurement turns "too slow" into a baseline (you can't show improvement without one), and the layout list tells you what the riskiest part of the build is. An FDE's first deliverable is usually understanding with a number in it, then the smallest build that tests the riskiest assumption.
This course trains six skills, one or two per module. You need all six, but not equally on every engagement.
| Skill | What it means at a customer |
|---|---|
| Customer discovery | Find the problem worth solving, who cares, how success is measured |
| Data wrangling | Profile, clean and reconcile data that's messier than promised |
| Integration engineering | Rate limits, pagination, auth, schema surprises |
| LLM application engineering | Prompts, structured outputs, tool use, retrieval, agents, evals |
| Production engineering | Deploy and debug in environments you don't control |
| Communication | Briefs, demos, feedback to your product team |
Traits that matter: bias to a working thing, owning the outcome rather than the ticket, and low ego about the stack (sometimes the answer is a SQL query and a cron job, not an agent).
Low ego about the stack deserves an example. Suppose Brightline's slowest step turns out to be sorting faxes by clinic, and every fax already carries the clinic's fax number in its header. A lookup table solves it. An FDE who reaches for Claude anyway has spent the customer's time and trust on the wrong thing. Use the model where judgment on language is needed, and plain code everywhere else.
Interviewers and customers both ask "so what does an FDE actually do?" Have an answer that fits in a minute and is specific:
I'm the engineer who owns a customer's outcome after the sale.
Week one: find the problem with a number in it, and the people who decide.
Weeks two to six: build the smallest thing that tests the riskiest assumption,
on their real data, and run it with a small pilot group.
Every week: measure against the baseline and send a short written update.
Throughout: turn what I built by hand into patterns my product team can ship.
The right side runs Python in your browser: a scratchpad on readings, graded by tests on exercises. Claude calls use an offline simulator of the official anthropic SDK, so your code also runs against the real API with pip install anthropic and an ANTHROPIC_API_KEY.
Each module follows one fictional customer through one phase of an engagement, from Brightline's kickoff to production. Readings explain the method, exercises make you build a piece of it, and field drills put you in a conversation with the customer contact.
Try it now: press Run in the editor on the right. The scratchpad rates your six skills; change the numbers to your honest self-assessment and run it again to see where to focus.
Cheat card
- An FDE owns a customer outcome after the sale and writes the production code to reach it, in an environment they don't control.
- Deployments are where the best product insight comes from: repeated custom work should become product.
- Six skills: discovery, data wrangling, integrations, LLM application engineering, production engineering, communication.
FDE loop: discover (a number) → build (riskiest assumption first) → deploy (pilot, shadow mode) → measure (vs baseline) → feed back (customer update + product pattern) → repeatInterview line "I'd rather show a measured baseline on the customer's real data in week one than an architecture diagram, because the baseline is what every later result gets compared to."