Human-AI Software Developer (Mid)

The work

You will write the application code around models rather than train the models themselves. That means prompt and context construction, retrieval, tool calls, evaluation harnesses, fallbacks, and the APIs and interfaces that carry all of it to a user. You will build the human checkpoints. Which outputs go straight through, which ones stop and wait for a person, what that person is shown so the decision takes seconds instead of minutes. This is product work as much as it is inference work. You will make outputs explainable after the fact. When a client or a user asks why the system said what it said, the logs, traces and stored context should answer it without you reconstructing anything from memory. You will carry your own code in production, including the pager. You will write the evals that tell you a change made things better, not just different, and you will accept latency and cost limits as real constraints rather than someone else's problem.

What we're looking for

Five to eight years writing software that other people depend on, with at least one system you took from design to production and then operated. Strong Python, and enough comfort in a typed language and in a frontend to build an end-to-end feature without waiting for someone else. Hands-on experience shipping something built on LLMs or similar models: retrieval, agents, structured extraction, classification, whatever the shape. We care that you have seen it fail in front of real users and know what you changed. SQL and a relational database you know well. Queues, caching, and the usual mechanics of keeping a service up. The habit of measuring. If you cannot say how you knew a change helped, the change did not happen.

How we hold the work

Every system here is held to four principles, in this order. Context before capability: we do not ship a feature because the model can do it, we ship it because we understand the situation it lands in. Wellbeing is the metric: the system is judged on whether the person on the other side is better off, not on engagement, retention, session length or task volume. Judgment stays human: consequential decisions have a person making them, and the software is built to make that possible rather than nominal. Explainable under pressure: an output has to be traceable and defensible on the day it is challenged, not just on the day it was designed. We work as Conscipact HAI. AI carries volume, humans carry judgment. Every output has a human accountable for it, including yours. The method is part of what we deliver to clients, so how you build is visible, not just what you build.

Where this work goes

Roughly half our engineering is our own products for human wellbeing. The other half is systems built for enterprise clients under the same standard. You should expect to move between the two, and to occasionally explain your design decisions directly to a client engineer or a client's risk function. We are not an AI consultancy with a values page, an ethics advisory firm, or a wellness brand. We ship production systems and then keep them running.

What we'll ask you

No surprises — these are the questions in the application, so you can think about them before you start.

• Describe an LLM-backed feature you shipped to real users. What was it, what did you own, and how did you measure whether it was working?
• Tell us about a time a model your code depended on produced a wrong or harmful output in production. How did you find out, and what did you change in the system afterwards?
• Give one example of a place where you deliberately kept a human in the loop instead of automating a decision, and how you decided where that line sat.
• Have you been on-call for a service you wrote?