By FDE Interview Prep, a SourceStrongAI company

In this post11 sections
  1. What employers ask for: projects, portfolios and demo videos
  2. What makes a project read as FDE work
  3. Template one: an assistant over a real document set
  4. Template two: an integration that joins two messy public sources
  5. Template three: an internal tool for a real user
  6. The README a hiring manager reads
  7. The demo script
  8. Putting the project on your résumé and in the interview
  9. Questions people ask
  10. Keep reading
  11. More from the blog

Your GitHub has a to-do app, a weather dashboard and three tutorial clones, and none of them has ever been used by anyone but you. Another one won’t make a recruiter stop scrolling. The portfolio project that does is small, runs on real and messy data, and has a named user who tried it, plus a README and a short demo that show the calls you made. This post is part of the FDE resume guide. It gives you three project templates, one AI, one data integration and one internal tool, and the README, demo script and interview words to go with them.

What employers ask for: projects, portfolios and demo videos

  • Palantir says to build. Its careers page on open-ended questions advises: “Find a public dataset, and think of something interesting to do with it.” Source 1Palantir Careers | Navigating Open-Ended QuestionsPublisherPalantir TechnologiesSource typecompany hiring page
  • Ramp names the project as a signal. Leo Mehr, Ramp’s director of engineering, said in June 2026 that a public technical project combined with people-facing work, such as a cashier or sales job, is a great example of someone who might be good for FDE. Source 2What is a Forward Deployed Engineer? (FDE Explained) feat. Leo Mehr of Ramp (YouTube auto-generated English captions)PublisherdearCC (Clara Shih), YouTubeSource typerecorded talk or interview
  • The best FDEs build compulsively. Shilpa Balaji, a former Palantir FDE recruiting lead, told First Round Review in February 2026 that the best FDEs she worked with were compulsive builders. Source 3So You Want to Hire a Forward Deployed EngineerPublisherFirst Round ReviewSource typenews report
  • Some applications ask outright. LangChain’s application form for its Deployed Engineer, Professional Services role in San Francisco requires a portfolio. Source 4Application formPublisherLangChain (Ashby job board)Source typecompany job posting
  • One track asks for a demo video. Palantir’s Build to Apply track, which is for its US Government role and not the application, asks applicants to build a Foundry and AIP project and submit a demo video, and says this can mean expedited consideration. Source 5Deployment Strategist, Build to Apply - US Government (Washington, D.C.) - Palantir on LeverPublisherPalantir Technologies (Lever)Source typecompany job posting Those postings ask for a video under 5 minutes, as an unlisted YouTube link. Source 6Deployment Strategist, Build to Apply - US GovernmentPublisherPalantir Technologies (Lever)Source typecompany job posting

We found no source that says a portfolio is required across FDE hiring, and none that says how a project is scored. What the sources do support is simpler: when you have built something real, it gives the recruiter a reason to call and the interviewer something concrete to ask about.

What makes a project read as FDE work

Palantir, which says it pioneered the role, describes it as engineers directly with customers. Source 7Forward Deployed Software Engineer (New York, NY) - Palantir on LeverPublisherPalantir Technologies (Lever)Source typecompany job posting Your project should be a small copy of that job. Hiring leaders describe the same profile from different angles:

  • Ramp’s engineering blog says what matters most in FDE recruiting is drive, communication and the will to solve customer problems end to end, not technical perfection. Source 8Forward Deployed Engineering (Leo Mehr, Director, Engineering)PublisherRamp Builders (engineering blog)Source typecompany blog
  • Bob McGrew, introduced as an early Palantir executive, said on the Lightcone podcast (Y Combinator, YouTube) in September 2025 that for Deltas, Palantir’s name for its forward deployed engineers, you want someone really good at prototyping, and that the wrong profile is a craftsman who loves getting the abstractions exactly right. Source 9The FDE Playbook for AI Startups with Bob McGrew (YouTube auto-generated English captions)PublisherY Combinator, YouTubeSource typerecorded talk or interview Source 10Palantir Careers | Students and Early TalentPublisherPalantir TechnologiesSource typecompany hiring page
  • Colin Jarvis of OpenAI, on a South Park Commons panel in May 2026, named as a failure pattern people who come to love the form of what they built more than its function, and said the best FDEs tear a solution up and build something different when users don’t use it. Source 11The Future of Forward Deployed Engineering | OpenAI, Ramp, Nominal, DatalandPublisherSouth Park Commons (YouTube)Source typerecorded talk or interview

From that, we use four tests. They are our method, not a rubric any employer publishes:

  1. A named user. A person with a job, not “users”. “The treasurer of my climbing club” passes. “Anyone who wants to track tasks” fails.
  2. A messy input. Real data with gaps, duplicates and odd formats, or real documents written by people who never imagined a parser.
  3. A first version someone used. Ugly is fine. Unused is not.
  4. A change you made because of them. The feature you cut, the screen you rebuilt, the thing you learned they actually needed.

The fourth test is the one a tutorial project can never pass, which is why it is the one to build toward. Here is how common projects look before and after:

Typical projectFDE-shaped version
A to-do appA shift-swap sheet the café team used for a month
A chatbot over “any PDF”An assistant over one city’s housing code, for a tenant group
A dashboard of a famous datasetInspection and license records joined for a local reporter
A clone of a popular appA reimbursement tracker the club treasurer still uses

If the customer part worries you more than the building part, how to show you can handle customers is the companion to this post. ElevenLabs’ FDE postings ask for experience working with customers, and say it is fine if that came from student clubs or side projects. Source 12Forward Deployed Engineer - Software Engineer - North America @ ElevenLabsPublisherElevenLabs (Ashby job board)Source typecompany job posting A project with a real user counts as that experience.

Template one: an assistant over a real document set

The user. Someone who answers the same questions from a pile of documents: a tenant association volunteer fielding questions about the housing code, a club secretary explaining the bylaws, a TA answering questions about the course policy.

The data. A public, bounded document set: a city’s municipal code, a school board’s meeting minutes, a university’s handbook. The messiness is the point. Scanned pages, tables, amendments that override older sections, and two documents that disagree.

The first version. A , meaning every layer, thin, joined end to end:

  • Load the documents, split them into passages and keep each passage’s source and section.
  • Retrieve the relevant passages for a question and answer only from them. This is retrieval-augmented generation.
  • Show the cited section under every answer.
  • Say “I can’t find that in these documents” when retrieval comes back weak, instead of guessing.

How you know it works. Sit with your user and write down the questions they get, together with the right answers and the sections they come from. That list is your golden set. Three rows might look like this:

QuestionWhat a pass looks like
How much notice before a rent increase?Cites the amended section, not the superseded text
Can my landlord enter without notice?Cites the entry section, with its exceptions
Is my building rent-stabilized?Refuses: not in these documents

An illustrative first run: 31 of 40 right, and 6 of the 9 misses quoted superseded sections. That line tells you what to fix.

Run the assistant against it before and after every change, and write down what it got wrong. The free question how do you know your AI system works? is the interview version of this step, and it is worth rehearsing with your own project as the example.

What to show. One good answer with its citation, one question it correctly refuses, and one failure you found and fixed, such as an amended section it used to quote in its old form.

Template two: an integration that joins two messy public sources

The user. Someone who needs two datasets to agree: a local reporter asking which restaurants approved for outdoor seating have poor inspection results, a food bank coordinator matching pantry sites to transit stops, a small-business association checking which members hold current permits.

The data. Two public sources that describe the same things in different ways. NYC Open Data publishes both halves of the reporter’s question: DOHMH New York City Restaurant Inspection Results and Open Restaurant Applications (Historical). They share no restaurant ID: inspections key on a CAMIS number, and applications carry names as the applicant typed them. Shared BBL and BIN codes narrow a match to a building, not a restaurant. Names never match cleanly: “JOE’S PIZZA, INC.” in one file is “Joe’s Pizza LLC” in the other.

The first version. Pull both sources, normalize the join key, join, and put everything that did not match into a review list instead of dropping it. The heart of it can be this small:

import re
TAIL = {"inc", "llc", "corp", "co", "ltd"}

def name_key(name: str) -> str:
    name = name.lower().replace("’", "'")
    name = name.replace("'", "").replace("&", " and ")
    words = re.sub(r"[^a-z0-9]+", " ", name).split()
    if words[:1] == ["the"]:
        words = words[1:]
    while words and words[-1] in TAIL:
        words.pop()
    return " ".join(words)

name_key("JOE'S PIZZA, INC.")  # joes pizza
name_key("Joe’s Pizza LLC")    # joes pizza

It only strips legal suffixes at the end, so “Co-op Grocery” survives as co op grocery. Then report what the key missed. “Joes Pizzeria” will not match “Joe’s Pizza”, and that is fine, as long as it lands in the review list with the address beside it and you can say how many records sit there. Hiding unmatched rows is the mistake that makes a data project untrustworthy.

The customer email. Finish the way a real engagement would. Ethyca’s published take-home for its forward deployed privacy engineer role has the candidate reproduce and diagnose a customer’s API error, then write the email reply to that customer. Source 13Ethyca Technical Challenge -- Forward Deployed Privacy Engineer (FDPE)PublisherEthyca (GitHub)Source typecompany website Borrow that step: write a short email to your user that says what matched, what didn’t, why, and what you need from them to close the gap. Put it in the repository. Our lesson on writing for customers covers the shape of that email.

What to show. The two raw files side by side, the join, the review list, and the email.

Template three: an internal tool for a real user

The user. A small group with a process that runs on a spreadsheet, a chat thread and someone’s memory: a club treasurer handling reimbursements, a TA running an office-hours queue, a family business tracking stock.

The first version. The smallest tool that replaces one step of that process. For the treasurer, that might be a form that takes a receipt photo and an amount, a list the treasurer approves from, and an export for the bank. No accounts system, no roles model, no settings page.

The part that makes it FDE work. Put it in front of them within days, not weeks, and watch them use it. Keep a short log:

  • what they asked for;
  • what they did instead of what you expected;
  • what you changed, and what you deleted.

That log is the most valuable file in the repository. It shows the habit Jarvis described: function over form. Source 11The Future of Forward Deployed Engineering | OpenAI, Ramp, Nominal, DatalandPublisherSouth Park Commons (YouTube)Source typerecorded talk or interview If your user ignored the dashboard you were proud of and only used the export, say so and show that you cut the dashboard.

Our walking skeletons lesson goes deeper on what a first version must include, and what you may fake in it.

The README a hiring manager reads

Assume the reader skims your README and never runs the code. Write it for that reader, under a title that names the user, such as # Housing-code assistant for the Elm Street tenants' association, in this order:

  1. Who it’s for: the association’s volunteer who answers housing-code questions from tenants.
  2. What it does: answers from the city code, with the section cited.
  3. What happened: what the user did with it, in their words.
  4. Decisions: one choice, the option you rejected, and why.
  5. Known limits: what breaks, and what you’d do next.
  6. How I used AI: what you generated, and what you checked by hand.
  7. Run it: one command.

A few rules make it work:

  • Open with the user, not the stack. The stack goes at the bottom, if anywhere.
  • Add a screenshot or a short GIF near the top, so a reader on a phone sees the product without cloning anything.
  • Make the decision section real. “It refuses when no passage scores above the threshold, because a wrong answer about eviction notice is worse than none” shows judgment. “Built with React” does not.
  • Write the limits yourself. An interviewer who finds a gap you already named learns you are careful. One who finds a gap you didn’t learns something else.

The demo script

Record a short screen demo even if no application asks for one, and link it from the top of the README. If an application sets a limit, follow it exactly: Palantir’s Build to Apply postings, for the Deployment Strategist role, ask for under 5 minutes. Source 6Deployment Strategist, Build to Apply - US GovernmentPublisherPalantir Technologies (Lever)Source typecompany job posting The timings below are our suggestion for a project like these.

0:00 The user and the pain. “This is for the volunteer who answers housing questions for a tenants’ association. Today they search a long PDF by hand.”

0:30 The live run. Do what the user does. Ask a real question, show the answer and the cited section.

1:30 The messy part. Show the failure you fixed, or the review list of unmatched records. “The first version quoted a section the city had amended. Now it checks the amendment table first.”

2:30 One decision. “I made it refuse when retrieval was weak, because a wrong answer about eviction notice is worse than no answer.”

3:00 What the user said and what’s next. End on the next step you’d take with this user, not on a feature list.

For recording mistakes and the words for trade-offs, the take-home walkthrough video post has a full script. If you are applying to Palantir’s Deployment Strategist role through Build to Apply, our Build to Apply post covers how to scope and film that demo.

Putting the project on your résumé and in the interview

On the résumé, give the project one line with a user and an outcome, and put the link beside it:

  • Weak: “Built a chatbot with Python and a vector database.”
  • Strong: “Built an assistant over the city housing code for a tenants’ association volunteer; they now use it to answer tenant questions, with the section cited. Link.”

A new graduate or career changer can move the project above experience, because that is where the evidence is. The FDE resume guide covers where it sits on the page.

In the interview, the project gives you the answer to a question you can prepare for today. The free question take me through a project from your résumé asks for the problem, your design, what broke and what you would change. Open like this:

“The user was the volunteer who answers housing questions for our tenants’ association. The problem was that the answers lived in a PDF nobody could search. I’ll walk you through the first version, the thing that broke when she used it, and what I changed.”

Then stop and let them steer. You have given them three places to dig: the user, the failure and the change. Have more detail ready for each.

To practice that conversation from the other side, the free practice case puts you in front of an AI customer with a vague problem on a ten-minute clock, and scores you with quotes from what you said.

Prepare a second version of the same story for a mixed room. Presenting a past project to engineers and executives is the drill for that.

Common mistakes, and the fix for each:

  • Five half-projects. Link one finished project. Archive the rest or leave them unpinned.
  • No user. Find one before you write more code. A friend’s club counts if they use it.
  • A stack-first README. Rewrite the first three lines around the user.
  • Hidden failures. Put them in “Known limits” and the demo. They are your best interview material.
  • Claiming more than you did. If a library or an assistant did the hard part, say so and talk about the decisions that were yours.

Pick one template this week, name your user by Friday, and put the first version in their hands before you polish anything. The mixed-audience drill, the walking skeletons and customer-email lessons behind these templates, and model answers to all 180 questions are in Pro, which starts with a 7-day free trial.

GlossaryForward deployed engineerA software engineer who builds and ships production systems inside a customer’s problem and environment, accountable to that customer’s outcome.More on Forward deployed engineerGlossaryDeployment strategistA customer-facing role that works out the customer’s questions and scope beside FDEs; at some employers the title means a product-manager or quota-carrying role instead.More on Deployment strategistGlossaryForward deployed software engineerPalantir’s title for its FDE role, called Delta internally; OpenAI and EY also post FDSE titles, each with its own duties.More on Forward deployed software engineerGlossaryEmbeddingA vector representation of text or other data used to measure similarity of meaning.More on EmbeddingGlossaryWalking skeletonThe thinnest end-to-end version of a system that performs one small real function across its main components, built first and then extended.More on Walking skeletonGlossaryRetrieval-augmented generationAnswering with a model that is given passages retrieved from a document collection as context.More on Retrieval-augmented generation

Questions people ask

Do FDE applications ask for a portfolio?

Some do. LangChain’s Deployed Engineer, Professional Services application form in San Francisco requires a portfolio, and Palantir’s Build to Apply track for its US Government Deployment Strategist role asks applicants to build a Foundry and AIP project and submit a demo video. Where none is asked for, a linked project still gives the recruiter and the interviewer something concrete to ask about.Source 4Application formPublisherLangChain (Ashby job board)Source typecompany job postingSource 5Deployment Strategist, Build to Apply - US Government (Washington, D.C.) - Palantir on LeverPublisherPalantir Technologies (Lever)Source typecompany job posting

What kind of project should I build for a forward deployed engineer role?

Build something a specific person could use, on real and messy data, such as an assistant over a public document set, an integration that joins two public sources, or a small internal tool for a club or team. Palantir’s own preparation advice for open-ended questions suggests finding a public dataset and thinking of something interesting to do with it.Source 1Palantir Careers | Navigating Open-Ended QuestionsPublisherPalantir TechnologiesSource typecompany hiring page

How long should an FDE project demo video be?

If the employer sets a limit, follow it. Palantir’s Build to Apply postings, which are for its US Government Deployment Strategist role rather than the FDSE application, ask for a demo video of under 5 minutes, as an unlisted YouTube link. Where there is no limit, keep it short enough to watch in one sitting, and open with the user’s problem rather than the stack.Source 6Deployment Strategist, Build to Apply - US GovernmentPublisherPalantir Technologies (Lever)Source typecompany job posting

Keep reading

Guides

Lessons

Questions

More in Preparation