Reasons, not just scores
A prediction without its reasons is a guess with a confidence interval. Every output Kaimei produces comes with the inputs that moved it, in the output's own units.
About Kaimei
Kaimei means to clarify, to elucidate, to reveal through investigation. We build the tools that let teams do exactly that with the machine-learning models they already run.
Why we exist
Machine learning now decides who gets a loan, which claim gets a second look and what a patient is told to watch for. Most of those models are black boxes, judged by an accuracy number and trusted until something goes wrong.
A model can be right for the wrong reasons. Two models can agree on an answer while believing opposite things about the world. A production feed can start sending a field in the wrong units and nothing in a test suite will notice, while every prediction quietly moves.
Kaimei exists so that every one of those models can answer the question the people affected by it are entitled to ask: why? We attach to the models teams already have, in the frameworks they already use, and give every prediction its reasons, every model its anatomy, and every deployment a watch.
kaimei · to clarify, to elucidate, to reveal through investigation
Same client, two models
Accuracy tells you what a model got right. It never tells you why. Here is one card client from a public credit dataset, scored by a random forest and by a logistic scorecard.
| Input | Random forest | Logistic scorecard |
|---|---|---|
| Worst status in six months = 3 months late | +13.8 | +14.1 |
| Previous repayment = 2 months late | +11.2 | +8.0 |
| Latest repayment = 1 month late | −1.2 | +8.1 |
| Share of previous bill paid = 0% | +1.1 | −5.1 |
| Last payment = NT$0 | +0.7 | +1.4 |
| Credit limit = NT$20,000 | +2.5 | +4.5 |
| Balance change over six months = −5% of limit | +1.9 | +0.6 |
| Statement balance = NT$18,648 | −0.6 | −0.7 |
| Utilisation = 93% | +0.1 | +0.4 |
| Age = 35 | −0.2 | −0.0 |
01The answer
Two models score the same card client: 35, a NT$20,000 limit with 93% of it used, one month behind on the latest statement. The random forest puts the chance of default at 54.6%. So does the logistic scorecard, and the client did default. A leaderboard stops here.
02The reasons
The same explanation, run on each model for the same input. Each bar is one input raising or lowering P(default), in percentage points, from that model’s average: 25.3% for the forest, 23.4% for the scorecard. The rows line up, so you can read straight across.
03Where they agree
Three months late at worst in the last six statements, two months late on the one before this. The two inputs overlap. Together they add 25.0 points in the forest and 22.1 in the scorecard: most of the distance from an average client, in both.
04Where they split
Set the reference clients’ latest statement to on time, one month late, then two. The forest’s average answer goes 19%, 21%, 42%: a step. The scorecard’s goes 20%, 27%, 34%: a slope. So one month late adds 8.1 points in the scorecard and takes 1.2 off in the forest, which weighs the month before more.
05Opposite pulls
The client paid none of the previous NT$18,061 statement. The forest adds 1.8 points for that. The scorecard adds 1.4 for the amount, then takes 5.1 off for the share. Scored as having paid in full, its answer rises to 66.1%. In the data, 52.5% of those who paid nothing defaulted; 13.9% of those who paid in full did.
06The what-if
Hold the history fixed: age, limit, past statements. Ask each model for the smallest change to this month’s statement that flips its answer. The scorecard: a latest statement with no balance, and 54.6% falls to 27.1%. The forest’s search finds no flip. Its one step, paying 7% of the last bill, reaches 53.3%.
Same answer. Different reasons. Only an explanation tells them apart.
The two models are 0.04 apart on AUC and reach the same decision for 95.4% of the 6,000 held-out clients. Only their explanations show that one counts every month late and gives credit for paying nothing, while the other reads one month late much like on time and barely moves for any single change to this month’s statement. Which of those you would defend to a customer or a supervisor is your call. Kaimei explains every model you upload the same way, so you can make it before one goes live.
Data: Default of Credit Card Clients, UCI Machine Learning Repository (Yeh & Lien, 2009): 30,000 card clients of a bank in Taiwan, April to September 2005. Both models were trained on 24,000 of them; client 1105 is one of the 6,000 held out, and the AUCs are measured on those. Attributions are exact KernelSHAP over 58 reference clients, as the Kaimei API returned them. The averages in step 4 are each model’s behaviour curve from its full analysis, and “paid in full” in step 5 is one call to the predict endpoint. The default rates are counted over the 2,220 clients in the data who had a previous bill and paid nothing, and the 6,823 who paid it in full. The what-ifs are the API’s counterfactual search with age, credit limit, previous repayment, worst status in six months and six-month balance change held fixed. That search is greedy. Checked directly, no single input of this month’s statement takes the forest below 50% (lowest 52.2%), though two moved together can (48.5%), and with the balance change left free its search does cross the line, at 49.2%.
What we believe
A prediction without its reasons is a guess with a confidence interval. Every output Kaimei produces comes with the inputs that moved it, in the output's own units.
When a model is small enough we enumerate rather than sample, and say so. When we approximate, the method and its limits are part of the response, not a footnote.
We never store your model file. Upload shapes only, or keep the model behind your own endpoint, and Kaimei still does its job without your weights leaving home.
An Apache-2.0 client, a plain REST API, worked examples on public datasets and tutorial notebooks that run end to end. If a capability is described on this site, you can reproduce it yourself.
How we work
Everyone who works on Kaimei writes code and reads the research. The person who answers your issue is usually the person who built the feature.
The worked examples on this site are ordinary models trained on six public datasets, from card default to city power demand. Every explanation, analysis and drift figure quoted from them is what the API returned. The step-by-step tutorial, on the Titanic passenger list, runs in public notebooks with its outputs committed.
We are a small team working on a hard, important problem. If that sounds like your kind of work, or you want Kaimei on your models, we would like to hear from you.