The founder’s note

I built the machine. Then I made it work for me.

Every service like this shows you its best case. So here is the only honest compliment I could pay my own product: I put my own interview through it, as a customer, and published whatever came out.

The role is real — Staff Applied AI Engineer on the AI Experiences & Agents team at Linktree. The inputs were one job ad and my LinkedIn profile, nothing else. It went through the same pipeline every customer’s does, including two rounds at the QA gate, and it was delivered in three hours and fourteen minutes. Every claim about the employer below links to the public source it came from. Nothing here is templated.

I finished reading it wanting the job more than when I started — which is either very good research or a founder losing his objectivity. You can judge that yourself, because the whole thing is below.

This is the exact file the machine produced. Yours takes 24 hours.

Prefer a government role? The first example dossier — a federal infrastructure PM interview. That candidate got the job.

The Front Foot Dossier · File FFP-002Prepared in 24 hours

Interview brief · Linktree

Staff Applied AI Engineer, AI Experiences & Agents.

Linktree publicly bet the company on staying small and letting AI buy its velocity. This team is the sharp end of that bet. Everything in this file serves one idea: walk in as the person who makes the bet pay — someone who ships AI product and drives the human adoption their CTO calls the hard part.

Prepared for
Adam Rappaport
Panel
TBC — expect the hiring manager plus senior engineers; verify names when the invite lands
Program
AI Agents & Experiences team — creator- and visitor-facing AI
Inputs
One job ad + one LinkedIn profile
Turnaround
3 hours 14 minutes from inputs to delivered (the promise is 24)
Evidence
5 public sources — every employer claim cited

Status

The founder's own file

Adam built the machine, then ran his own interview through it as a real customer — two QA rounds, delivered in 3h 14m. This is the unedited result.

Get your dossier — $99 AUD

$99 AUD founding rate · $199 AUD standard after launch

Jump to the mock interviewTap any tile or term for detail · sources open in a new tab

Section 01The brief

They bet the company on your job description

The one thing to internalise before anything else in this file.

In 2025 Linktree made a deliberate, public bet: hold the company under 200 people — about 190 today, roughly 100 of them engineers — and buy product velocity with aggressive AI adoption instead of headcount. CTO Farnaz Azmoodeh: “we bet on being able to drive more impact by leaning into AI more aggressively.” The AI Experiences & Agents team is the sharpest end of that bet — the people who turn frontier-model capability into shipped product. Source: First Round Review — Scale Smarter, Not Bigger: How Linktree is Using AI.

Every answer you give should read as evidence the bet pays: ship fast, evaluate deliberately, keep the system simple, and bring the humans along.

Section 02The program

What AI Experiences & Agents actually ships

The actual work you would be hired to deliver.

The JD's own frame: turn frontier model capabilities into useful, trustworthy product experiences for creators, businesses and their audiences. It is explicit that this is a build-and-ship role, not research — concept to prototype to production, multi-agent workflows combining prompting, tool use, retrieval, memory and orchestration, eval suites and regression harnesses, and optimisation across latency, reliability, cost and safety. Source: LinkedIn — Staff Applied AI Engineer, AI Experiences & Agents (Linktree job ad).

The strategic direction is social commerce. Product leadership has said it plainly: “there are countless ways AI will positively impact our Linkers, and one we are laser-focused on is social commerce” — product suggestions, smart calls-to-action, AI-assisted link placement, onboarding cut from minutes to seconds. Assume the roadmap bends toward helping creators sell. Source: SmartCompany — Linktree explains new AI play “laser-focused” on social commerce.

Section 03Why you fit

Their seven asks, your evidence

The JD's “Who we're looking for” list, answered line by line from your own history.

  • Shipping production systems end-to-end. Three decades of it — from building and deploying an early NLP auto-response system for AT&T Wireless at Kana (2000–02), to shipping a community feature for the oldnavy.com launch, to productising M365 Optimiser at Data#3. Different eras, same discipline: it isn't done until it's live.
  • Hands-on LLM products and agents. Your declared top skills are Enterprise GenAI Adoption, Azure AI Foundry and Multi-Agent Framework Development — daily hands-on with Copilot Agents, Copilot Studio and Foundry across live enterprise deployments at Data#3.
  • Evals for AI output quality. You trained and tuned a Bayesian email-classification network in 2001 — label, measure, retrain, redeploy. That IS the eval loop, twenty-five years early. Frame modern eval suites as the same discipline with vastly better tooling, then show you've designed one for THEIR product (see the mock interview).
  • Strong product sense. Your Data#3 role is equal parts product dev, GTM, sales and evangelism — a job that only works if you can find the real user problem inside enterprise ambiguity and cut it down to what ships.
  • Simple-architecture trade-offs. Gap's internal innovations team was literally the buy/build desk. Thirty years of watching complexity die gives you the scar tissue to argue for the simplest system that ships — which the JD names as a value.
  • Judgment on safety, trust and abuse. Tenable certifications and years supporting Data#3's security practice — Purview, Entra, Sentinel, Defender. You speak security natively, which most applied-AI engineers cannot.
  • Communicates and collaborates cross-functionally. You present a daily GenAI news briefing on camera, enable sellers and customers for a living, and founded Vibe Academy to teach beginners to ship. Explaining trade-offs clearly is your default register.

Section 04The numbers

Numbers to carry into the room

Tap any card for the detail and its source.

Section 05The role

What you'd own, in their words

The JD's six deliverables, decoded.

Ship creator- and visitor-facing AI end-to-endPrototype through production. The team's identity is build-and-ship — arrive with opinions about what to ship first (three ready-made ones below).
Design multi-agent workflowsPrompting, tool use, retrieval, memory/state, orchestration, HITL. Your Multi-Agent Framework Development skill, in their vocabulary.
Create eval suites and regression harnesses“So AI quality improves deliberately rather than by anecdote” — the JD's sharpest line. Have an eval design ready for one of their features.
Translate ambiguous user problems with product/design/data/trustCross-functional by default — and remember their workflow hands you prompts as specs, not tickets.
Optimise latency, reliability, cost and safetyFast, dependable, viable at scale. Their moderation 25× story is the internal benchmark for this kind of win.
Turn production feedback into model, prompt and product improvementsSupport signals and behavioural data close the loop — your propensity-analysis history is this exact motion.

Notice what's absent: model training. This is applied engineering — orchestration, evals and product judgment, not research. The JD says so explicitly: LinkedIn — Staff Applied AI Engineer, AI Experiences & Agents (Linktree job ad).

Section 06The patch

Answer the doubt before they ask it

Two honest gaps sit between your CV and this JD. Both are survivable — if you raise them first.

“Your title says Sales Specialist — this is an engineering role.”

The move: own it in the first minute, then reframe the seat as chosen. Say this: “Fair — and deliberate. I took the adoption seat because that's where enterprise AI actually fails, and your CTO says the same thing: the hard part is human habits, not models. I've spent eighteen months running the playbook she runs — enablement, usage-driving, agents into production — inside a 1,200-person integrator. The difference is I also build: Copilot Studio agents live in customer environments, Foundry patterns, multi-agent frameworks. Sales title, builder's hands. Ask me anything technical in this interview and let the answer be the evidence.” Then stop.

“When did you last write production code at a product company?”

The move: receipts, not reassurance — and volunteer a proof mechanism before they ask. Say this: “Not at a product company recently — at customer sites and in my own shipped projects, constantly. I'd rather show than claim: give me a take-home or a paid work-sample week and judge the code. And I'll note your own posting says don't rule yourself out over unticked boxes — your best AI adopters are senior engineers who treat agents like delegation. Thirty years of coaching juniors is exactly that muscle.” If they push on stack: name yours honestly, then ask what theirs is and close the gap out loud.

“Walk us through how you'd eval an AI feature.” (they WILL ask)

The move: never answer in general — answer for THEIR live feature. Say this: “For your profile-restyle: four rubrics — niche relevance, brand coherence, accessibility, suggestion diversity. A golden set of a couple hundred consented profiles across creator categories. An LLM judge scores the rubrics, calibrated monthly against human raters. It runs in CI, so a prompt change that drops a rubric blocks the deploy. Online, acceptance rate feeds the set. And I'd run the judge on a small model — you cut moderation costs 25× that way, evals deserve the same economics.” The full version is in the section above — rehearse it to 90 seconds. Kana 2001 is your lineage story, not your answer.

The claims these scripts lean on are cited in this file: the CTO's adoption problem and senior-adopter pattern — First Round Review — Scale Smarter, Not Bigger: How Linktree is Using AI — and the “don't rule yourself out” line and role scope — LinkedIn — Staff Applied AI Engineer, AI Experiences & Agents (Linktree job ad).

Section 07Vocabulary

Speak their language

Tap any term. Use these words naturally — never all at once.

Section 08Your stories

Three STAR answers, ready to load

Grounded only in your profile — drop in the numbers LinkedIn truncated before you walk in.

Section 09Mock interview

The questions they'll actually ask

Six likely questions, what each is really assessing, and how to structure the answer.

Q1Tell us about an AI-powered product you've built end to end. What broke, and what did you change?+

AssessingWhether the hands-on claim is real. They want a build with failure modes, not a demo story.

  • Pick one Copilot Studio agent deployment. Name the stack plainly: Studio, Foundry, the connectors.
  • Structure as build → what broke in real usage → how you measured it → what changed.
  • Close with the measured outcome — this team's culture is allergic to anecdote.
Q2How would you design evals for our AI profile-restyle recommendations?+

AssessingEval literacy against a real feature of theirs. The JD's sharpest requirement.

  • Define the quality dimensions first: relevance to the creator's niche, brand-safety, visual coherence, diversity of suggestions.
  • Golden set from real (consented) profiles across creator categories; rubric scoring via LLM judge, calibrated against human raters.
  • Regression harness gating deploys; online signal — restyle acceptance rate — feeding the golden set back.
  • Say the quiet part: evals are how quality improves deliberately rather than by anecdote — their own words.
Q3A creator-facing agent is too slow and too expensive. Walk us through what you'd do.+

AssessingProduction instincts under the latency/cost/quality triangle.

  • Measure before touching anything — per-step traces, cost per interaction.
  • Route to the smallest model that clears the quality bar, escalate on low confidence; cache and precompute what repeats.
  • Stream the UX so perceived latency drops even where real latency can't.
  • Cite their own precedent: moderation to Gemini Flash, 25× cheaper. Small-model routing IS the house style.
Q4How do you decide what NOT to build?+

AssessingStaff-level judgment; their stated value of keeping architecture simple.

  • Start from the user problem, not the capability — most agent ideas die when you ask what the creator actually needs.
  • Simplest system that ships: a well-prompted single call beats an agent graph you can't debug.
  • Give one real example of killing your own complexity — the scar earns the principle.
Q5Creators' audiences include minors and vulnerable people. How do you think about failure modes?+

AssessingJudgment on safety, trust, privacy and abuse — a JD bullet verbatim.

  • Threat-model by surface: creator-facing suggestions fail differently from visitor-facing agents.
  • Layered safeguards: policy prompts, output filtering, rate limits, HITL for anything that publishes.
  • Abuse signals feed the eval set — safety regressions are regressions, gated like any other.
  • Your security fluency (Purview, Entra, Sentinel work; Tenable certs) is a differentiator here — use it.
Q6Our PMs and designers prototype in Lovable and Figma Make, then hand engineers the prompt as the spec. How does that land?+

AssessingCulture fit to their prototype-to-prompt workflow — and whether “staff” means gatekeeper to you.

  • Embrace it: “the new PRD isn't the prototype, it's the prompt” — you know the line and you like it.
  • Your job is hardening: take the prompt-spec, add evals, observability, safeguards, cost ceilings.
  • One caution, offered constructively: consolidated prompts hide decisions — you re-surface them as acceptance criteria.

Section 10Ask them

Questions that make them sit up

Pick two or three. Each signals you've done the homework.

  1. 01What would this role need to deliver in the first twelve months for you to call the hire a success?
  2. 02How do evals gate shipping today — and where is quality still being decided by anecdote?
  3. 03How does this team's roadmap connect to the social-commerce focus — is the agent work ultimately in service of Store Link?
  4. 04Given the sub-200 bet, where does a Staff engineer here find leverage — the Devin fleet, internal tooling, or something I haven't read about?

Section 11Opening move

Your first sixty seconds

When they say “tell us about yourself” — this, then stop.

“You've made a company-level bet — stay under two hundred people and let AI buy the velocity instead of headcount. Making that bet pay is literally my day job: I build the AI practice inside a major Australian integrator — agents in production, and the human adoption work your CTO calls the hard part — and I've been shipping AI systems since I trained a Bayesian classifier for AT&T Wireless in 2001. I'm here because this team is the sharp end of the same bet.”

Then stop talking. Whatever thread they pull — adoption, agents, evals, the 2001 story — lands in your fit table or a STAR. The silence after a strong open is theirs to fill, not yours.

Section 12Sources

Every claim, cited

Five public sources. Read First Round's piece in full — it is the hiring manager's worldview in essay form.

End of file · FFP-001 · The outcome

“I walked in feeling like I'd already worked there for a month. Every line of questioning, we'd anticipated — I had their own review's language, my projects mapped straight onto their program, and answers ready in my own words. It's the most prepared I've ever been for an interview, and it showed. I smashed it — and I got the job.”
— G — senior infrastructure PM, the first Front Foot Prep candidate

This dossier was built from one job ad and one CV in 24 hours. Yours is next.

Get your dossier — $99 AUDSee how it works

$99 AUD founding rate · $199 AUD standard after launch · complete dossier delivered within 24 hours of receiving the required inputs · every claim cited · doesn’t hold up? Full refund.