AI Engineer (Full-Stack)

Shortlists · Permanent · Europe, CET plus or minus two hours · Barcelona strongly preferred

Why this role exists

Shortlists is an AI-native recruitment CRM and ATS for recruitment agencies. We are building the assistant recruiters actually want: opinionated, does the work, does not just store records.

The AI is not a feature at the edge of this product. It is the product. So this is an AI engineering role as much as a full-stack one, and the two are not separable here.

The product is ahead of the team, which is a good problem and a temporary one. So this is not a role where you pick tickets off a queue. It is a role where you take a surface, own it, and make the calls on it. If you are the sort of engineer who has been waiting for permission to do that, this is the job.

Where we are

We are a Bifrost Forge company with a fast-growing customer base in Northern Europe. We are adding new customers all the time, so we are growing the team.

That means there is real product and real customers to work on, and more ambition than there are people, so what you own will be larger than your title suggests. As the team grows around you, you will have had the most context of anyone in it.

What you will do

• Own one or two product surfaces end to end: the UI, the Postgres schema underneath, the edge functions in between, the deploy, and what happens after it.
• Build and operate the LLM-backed parts of the product: prompts, retrieval, evals, cost and latency, and validating output you cannot fully trust.
• Turn a customer objection or a call transcript into shipped product, making the judgement calls in the middle yourself.
• Build with agents as your normal working method, and take responsibility for what they produce.
• Make the codebase easier for the next person and the next agent: sharper rules, better skills, gates that catch problems before a human reads the PR.
• Talk to the people using this. Our customers are recruiters who will tell you exactly what is broken.

How we work

• Small team, fast decisions. You will be in the room where they get made.
• Linear tracks reality, moved as work happens, not batched at the end.
• Feature branches, PR review, squash merge. main is protected and deploys on merge.
• Agents draft, humans approve, the system logs it.
• Async by default, with regular time together in person.

Our stack

TypeScript and React on the front end, Postgres and Supabase behind it, Next.js on the marketing side, deployed on Vercel, tracked in Linear. Multi-tenant, multi-language, with LLMs doing real work in the product rather than sitting in a sidebar. Node and Python are where we are heading as the AI services grow, so bring either and we will use it.

Coding agents are part of the repo, not bolted onto it. Claude Code and Codex are in our daily loop, and you will extend that setup rather than just use it.

You are not expected to have used all of this, and we will go through the details properly when we talk. Two things we do gate on: strong TypeScript and React fundamentals, and having written and shipped real code yourself, not only prompted for it.

Who you are

We are not filtering on years. We have hired people early in their career who ship more than people twice their experience, and the difference was always hunger.

• You have shipped full-stack features and operated them: migrations, auth, deploys, and the 2am production debugging.
• You have shipped something with an LLM in the critical path and kept it working. You know what breaks, what it costs, and how you tell whether it is getting better.
• You have worked inside a real engineering pipeline, in a team that had standards: branches, review, CI, releases. You know what good looks like and you will not need it invented around you.
• You build with coding agents every day, and you can explain how you prompt, review, and validate what comes back. You delegate the volume and keep the judgement on architecture, data, and security.
• You can read and reason about code you did not write. When an agent produces something plausible and wrong, you catch it.
• You notice when something is off and you fix it, rather than filing it.
• You do not need to be told what is next. And when something is genuinely missing, a key, a decision, an account, you ask immediately instead of quietly working around it.
• You want more responsibility than you have been given so far. That is the actual offer here.

Bonus points

• Security: auth, tenant isolation, secrets, dependency hygiene, GDPR.
• Supabase in production, especially RLS and edge functions.
• Real Postgres depth.
• js, Node, or Python.
• Multi-tenant B2B SaaS. Recruitment or HR tech.
• You have built your own agents, skills, or tooling.

How we hire

Four stages, usually about three weeks.

1. A 30-minute call.
2. A technical conversation where you walk us through something you built and we read code together.
3. A case. A real problem from our world: a customer objection and a slice of the product. How would you approach it, what would you build first, and what would you refuse to build. We discuss it together, we are not marking an exam.
4. A conversation with our CEO and our product lead, references, and an offer.

Use Claude Code, Codex, Cursor, whatever you actually work with. We are not testing whether you can code without AI. We want to see how you scope a problem, how you review what the agent gives you, what you reject, and what you are willing to ship.

Practicalities

Compensation is competitive and will include equity.

Our company language is English. You need an existing right to work in the EU or EEA. If you are outside that, tell us in your application and we will be straight with you about what we can and cannot support.

How to apply

Send us a CV or a profile, and one paragraph on something you built and owned: what you decided, what you got wrong, and what you would do differently. If some of it was built with agents, say so and say what you changed.

We reply to everyone within ten working days.

We use AI tools to help review applications. They assist the decision, they do not make it. A human reads every application.