AI explainability for critical decisions
Dakma explains the decisions your models make, from classic ML to LLMs, without ever seeing the model or its training data. Reveal the drivers, examine uncertainty, and keep people in the loop where it matters most.
BLACK-BOX BY DESIGN. NO MODEL ACCESS REQUIRED.
dakma-sdkpip install dakma-sdk
· Free
From output to understanding
Send a model's input and answer to the Dakma API and get back an explainability status, the checks behind it, and a plain-language reading your reviewers can act on.
See which parts of a prompt and its context shaped a response, using the same black-box approach and the same API.
Customer disputes a $2,400 charge at an electronics store in another country. The same card was used at home two hours earlier. Account open eight years.
Illustrative product view. Synthetic data, sample models and explanations shown for demonstration only.
How it works · two separate ways in
Pick the one that fits. Build on an open-weight model with the free SDK, or send any decision to the Dakma API. The two work independently.
Building on an open-weight model? The SDK runs next to your model in your own environment. It works on its own, with no connection to the Dakma API.
pip install dakma-sdkChatGPT, Claude, a vendor's system or a model you can't see inside. The API doesn't need to know which model it is. It works only from the input and the answer.
Need a paper trail? The governance add-on keeps explanations together with reviewer notes.
A clearer view of your AI
Local and global explanations show which features shape predictions, and how their influence differs across cases.
Feature attribution · classic MLView estimates with their limits. Spot ambiguous cases and give reviewers the context to ask better questions.
Estimate ranges · review signalsSee which parts of a prompt and its context shaped a response, with the same black-box approach and the same API.
Prompt & context attributionAn optional add-on that keeps explanations, the original input and answer, and review notes together in a record people can revisit.
Governance add-on · traceabilityFree SDK. Paid API.
Two separate products. The SDK is free for open-weight models you build and run yourself. The Dakma API works with any model and is the paid product. Governance is an optional add-on.
For open-weight models you build and run. Works on its own in your environment.
$ pip install dakma-sdk
View on PyPI →
For any model, including ChatGPT, Claude or one you can't see inside. It never needs to know which model it is.
For teams that need to show their work. Sits on top of the API.
Where clarity matters most
We focus on high-stakes decisions, where an unexplained "no" isn't acceptable to the person affected, the reviewer, or the regulator. Illustrative applications, not customer deployments.
Explain card-fraud and credit decisions, compare similar cases, and give review teams a clear basis for an approval or a decline.
Fraud teams · credit risk reviewersExplain admissions and scholarship screening models so reviewers can check that each decision rests on the right factors.
Admissions teams · review boardsExplore the factors behind predictive signals and make uncertainty visible to the domain experts evaluating an output.
Data scientists · domain expertsGovernance, as an add-on
Explanations are the core of Dakma. For teams that need a paper trail, the optional governance add-on keeps the evidence and human judgment together.
Explanations alone do not establish fairness, safety or regulatory compliance. They belong alongside validation, domain expertise, organizational policies and responsible human oversight.
A few useful answers
A grounded look at what explainability can and cannot do.
No. The Dakma API works only from the input and the model's answer, and it doesn't need to know which model you use. The SDK runs in your own environment and isn't connected to the API. Either way, your model weights, code and training data stay with you.
The Dakma API works with any model, including ChatGPT, Claude, vendor systems and models you can't see inside, because it never needs to know which model produced the answer. If you build on an open-weight model, you can use the free SDK in your own environment instead.
dakma-sdk is free on PyPI for open-weight models you run yourself. The Dakma API, which works with any model, is the paid product. The two are separate. Governance features are an optional add-on.
It helps people examine how a model's inputs relate to its outputs. Different methods reveal different views of that behavior; no single explanation tells the whole story.
No. Explainability can support investigation, but fairness, safety and compliance need their own evaluations, domain context and human oversight.
Model builders, domain experts, risk teams and decision reviewers. The goal is to preserve the technical evidence while making its meaning easier to discuss.
A clearer view starts here
Start free with the SDK, then switch on the Dakma API when you're ready to explain real decisions.