← Forward Deployed Engineer

ironco: Forward Deployed Engineer project

v1.02 Oct 2026Forward Deployed EngineerHiring step 2
Download the case folderfde-case-study-1.0.zip

The setup

ironco runs growth operations on TikTok Shop. We're an AI-first team: every person runs agents, and the agents run on a shared company brain, a folder of markdown files the team and the agents both read and write.

The FDE owns the agents the pods run. You build them, you prove they work, and you ship them to the people who use them. When a pod stops using an agent, that is yours to fix.

This case study is that loop, on made-up data: two blocks, three questions each, going from easy to hard. It starts where the job starts: with a ticket on your desk.

The rules

  • This should take about two hours. Spend as long as you want beyond that. Please no more than 6 hours. Tell us roughly how long you spent.
  • Use AI as much as you can. We expect AI to write most of your code. You own every line of it.
  • Use the stack you are fastest in, and any model you have access to. Put the install steps and the commands to run in README.md. Keys stay in your environment: none in the folder.
  • We grade what runs. We will run your commands on a transcript you have not seen.
  • Block A builds as it goes, so do it in order. Block B stands on its own. Skip what you must, and tell us which you skipped.
  • Keep a short log as you go in log.md: what you asked the AI, what it got wrong, and what you changed. Five to ten lines is fine.
  • All names, brands and numbers in this folder are fictional.

What's in the folder

  • brain/: a small company brain for a fictional creator agency, Brightline. Start with RESOLVER.md (where things go) and SANITIZER.md (what never goes in). It also has people pages, account pages, past meeting summaries, a reference sheet and NOW.md (this week's priorities).
  • inbox/: two raw call transcripts, call-1.md and call-2.md, not processed yet.
  • workers/skills/skill-recap.md: the skill file our recap agent follows.
  • workers/output/: bad-recap.md, a recap the agent wrote last week that a pod lead complained about, and bad-recap-source.md, the transcript it was written from.
  • adoption.csv: this month's dashboard for the recap step, by pod.
  • calls.csv: the per-call log the dashboard is built from.
  • slack-thread.txt: a short thread from one of the pod leads.
  • fde-ticket.md: the ticket operations wrote for you after last week.

A recap here always means two things: a summary of the meeting, and a list of action items with an owner and a due date for each.

The number the pods are measured on is the share of client calls that get a recap posted within an hour. The target is 80% for every pod by the end of next month.

Block A: Build and prove the agent

  • A1 (easy). Write one command that recaps a transcript into brain/knowledge/meetings/, following the brain's rules. Each run is a fresh AI session that sees only the skill, the brain and the transcript. Run it on inbox/call-1.md and inbox/call-2.md, and leave the recaps in the folder.
  • A2 (medium). Some of the brain's rules are too important to leave to a prompt, including what the agent may write and where. Decide which, and enforce them in code, before or after the model. In design.md, list each rule, where you enforce it, and why.
  • A3 (hard). Write checks that fail bad-recap.md and pass a good recap. Run your agent five times on each of the three transcripts in the folder and report the pass rate for each check in eval.md. Then edit skill-recap.md, rerun, and show before and after. Would one run have been enough to trust the change?

Block B: Ship it to the pod

  • B1 (easy). Rebuild adoption.csv from calls.csv in code. Say what the one in the folder gets wrong.
  • B2 (medium). One pod has stopped using the agent. Work out which and why, using anything in the folder. Add the fix to design.md, and build the smallest piece of it that runs.
  • B3 (hard). Answer fde-ticket.md in reply.md: what you ship this week, what waits, what you push back on, and what you need from the pod. Under 200 words.

Show your AI work

We expect AI to write almost everything here, and we expect you to own whether it is right. Add ai-notes.md, ten lines or fewer:

  • Setup. The tools and models you used, and what you pointed them at.
  • Split. What the AI wrote, and what you decided or changed yourself.
  • One catch. Something the AI got wrong. How you caught it, and whether you fixed the output or fixed the prompt, skill or check so it can't happen again.
  • One check. One piece of AI-written code you verified. How do you know it is right?
  • One limit. One thing in this folder you would not hand to an AI, and why.

Using more AI is only better if the output is right. We grade whether we can trust what you shipped, and whether it would still work tomorrow on a call you have never seen.

Extra credit (optional)

One pod's calls never reach the agent at all. Suggest a way in and build an alpha of it. Rough is fine, but it should run. Put it in extra/, with a README of five lines or fewer: what it does, how to run it, and what you would build next.

What to send back

  • This whole folder, zipped: your code and README.md, the two recaps, design.md, eval.md, reply.md, log.md, ai-notes.md, and extra/ if you did it.
  • How long you spent, and which questions you skipped.
  • Three sentences: what you would build next, and why.

Then book a 20-minute working session: tidycal.com/jeremycarr. It is not a presentation. We will run your commands on a transcript you have not seen, and we will ask you to change something live with your setup.