What the platform is designed to do.
Inverse AI is pre-product. This page describes the capabilities we are building, grouped by layer, and marks how far each one has come.
Explanations you can check.
Most explanation tools stop at the picture. Inspect scores it: does the credit fall where evidence can exist, and do different methods agree?
Mask-aware explanation scoring
PublishedScore Grad-CAM, SHAP and other attribution maps on whether credit falls inside the region that can carry evidence, such as the retinal disc, and not on borders or background.
Method comparison per decision
PublishedRun several explainers on the same prediction and show where they agree, where they split, and which grades or classes each handles better.
Lesion-level validation
In researchScore explanations against lesion outlines where datasets provide them, so a heatmap is tested against the pathology and not only the anatomy.
Decision report
In designOne page per prediction with the output, the explanations, their scores, and a plain statement of what was and wasn’t measured.
Change behaviour without retraining.
Representation steering edits what a running model does while its weights stay fixed. Intervene makes each edit legible, comparable, and costed.
Answer, clarify, or refuse
PrototypeA learned controller sits around an unchanged language model and decides, step by step, whether to reply, check with a second safety model, or refuse, and whether to send a draft once it is written.
STEER intervention cards
In researchEvery steering change is written up the same way: Signal, Target site, Edit, Execution, Result. Two interventions can then be compared line by line.
Strong-baseline evaluation
In researchEach intervention is measured against the untouched model and against a strong alternative such as prompting or fine-tuning. Beating a weak baseline doesn’t count.
Cost accounting
In researchReport what the change cost in general capability, compute, and latency, next to what it gained.
Text and multimodal coverage
In researchThe same standard applies to models that read and models that read and see. A result tested on one and not the other is labelled that way.
Keep the evidence in production.
A model that behaved in the lab can drift once it’s live. Monitor carries the explanations and interventions into deployment, building on the founder’s published MLOps work.
Model CI/CD
In designVersioned data, models, and pipelines, with continuous training and deployment. Every release can be traced and reproduced.
Explanation drift
PlannedAlert when the reasons behind predictions shift, not only when accuracy drops.
Fairness and subgroup checks
PlannedTrack performance and explanation quality across patient or user groups, release over release.
Audit trail
PlannedLog every model, explanation, and intervention for review against frameworks such as the EU AI Act.
Have a model that needs this?
We’re looking for design partners in healthcare, AI safety, and regulated industries.