CareersProduct

Product Manager

Decide what Kaimei builds next for the teams who run models and the people who have to trust them, then see each release through until a user can call it, read it and act on it.

How to apply The product problems

The role

Kaimei explains and monitors the machine-learning models teams already run. It reaches them three ways: a REST API, a Python client that mirrors it one for one, and a dashboard that draws attributions, insights, drift and a model’s anatomy. The people who use it range from the engineer who trained the model to the reviewer who has to sign it off, and they rarely want the same thing from an explanation.

As Product Manager you will decide what Kaimei builds next and why, and see it through to the moment a user can rely on it. You will talk to the teams running models in production, turn what they tell you into problems precise enough to build against, and hold each release to the standard we hold ourselves to: shipped, explained and shown running on public data.

Kaimei is a small team, and you will be its most demanding user. You will run every notebook, read every response, and file the issue when the docs and the product disagree.

What you will do

Own the roadmap across the API, the Python client and the dashboard, written down with the reason each item exists and what it displaces.
Talk to the teams who run models in production, from machine-learning engineers to model-risk reviewers, and turn what they ask for into problems the team can build against.
Write the spec before the code: who it is for, what the response contains, and how we will know it worked.
Shorten the path from sign-up to a first explanation of a user’s own model, and measure every step of it.
Work with researchers and engineers so a new method reaches users as something they can read: an attribution bar, a heatmap, an alert or a sentence.
Help write what ships: the docs, the changelog and public notebooks with their outputs committed.

What you would work on

Six product problems, each starting from something that ships.

Every capability below is live and documented. What is still open is how it should work for the people who use it. Pick one, or bring your own.

The first ten minutes

Ships today

A five-step quickstart goes from an API key to a model explained and analysed. It uses one of five built-in reference models, so there is nothing to deploy, and each Titanic notebook runs in about a minute.

quickstartbuiltinnotebooks

Open question

How quickly a team gets from sign-up to an explanation of its own model, and which step loses the most people on the way.

Quickstart

Every framework, one way in

Ships today

The client uploads scikit-learn models, Keras networks and PyTorch CNNs, which then run inside Kaimei. XGBoost, LightGBM, CatBoost, ONNX and Hugging Face models connect through a scoring endpoint, or bring the attributions they already compute.

uploadremoteattributions

Open question

Which framework should run inside Kaimei next, and how to know it is the right one before it is built.

Supported models

For the person who signs it off

Ships today

Every insight has a title, a one-paragraph summary, a severity and the numbers behind it. Attributions are in the output’s own units, and image explanations end with a plain-English summary and an evidence check.

summaryseverityevidence

Open question

What a model-risk reviewer, an auditor or a declined applicant needs to see before trusting an explanation, and in what order.

Insights

Alerts worth acting on

Ships today

Medium and high insights surface as alerts in the organisation and project overviews. Drift turns significant above a population stability index of 0.25, and robustness turns high above a 10% decision flip rate.

alertsdriftrobustness

Open question

Which thresholds a team should set for itself, where an alert should reach them, and how to keep alerts rare enough to trust.

Monitor production traffic

From one model to fifty

Ships today

Projects group models with roll-ups of traffic, alerts and the frameworks in use, and a feature matrix of the inputs each model relies on. Versions share a lineage, keep their own history and compare insights side by side.

projectsversionslineage

Open question

What a head of machine learning needs to see across every model on a Monday morning, and what they would do about it.

Projects

Trust before the first upload

Ships today

Kaimei never stores training data or the model file. Private weights upload shapes and statistics only, a scoring endpoint keeps the model on your own infrastructure, and every record is scoped to one organisation.

weights="summary"remoteorganization

Open question

What a security review asks before a team uploads anything, and how the product answers it before anyone has to ask.

What Kaimei keeps

You might be a good fit if you

have managed a developer product, an API or a machine-learning platform, and can point to something you shipped and what it changed;
understand how models are trained, deployed and monitored well enough to follow an engineer’s trade-offs and push back on them;
can read an API reference, run a notebook and write enough Python or SQL to answer your own questions;
decide with evidence from usage, interviews and the support queue, and can say what would change your mind;
write specs people want to read, and can explain a hard idea in plain language. It is most of what explainability is.

Strong candidates may also have

Experience as a machine-learning engineer, data scientist or model-risk analyst before moving into product.
Product work in a regulated setting, such as credit, insurance or healthcare, where a model’s decisions have to be explained to someone outside the team.
Experience with developer onboarding, documentation or usage-based pricing.
Hands-on use of SHAP, LIME or drift monitoring, from the user’s side.
Time as an early product hire at a small company.

How to apply

Send us three things through the contact form.

01
Your workA CV, or a link that does the job better: LinkedIn, a personal site or something you wrote.
02
One thing you shippedA product or feature you owned, the decision you are proudest of, and one you would make differently now.
03
One product problemA paragraph on one of the six problems above: who it matters to most, what you would ship first, and how you would know it worked.

Apply for this role Other roles

Use the product, then write to us.

The free plan covers the whole Titanic walkthrough, and every capability on this page is documented. Come with opinions about what to build next.

Start free Titanic example