Inverse AI
Research

Research is the foundation, not the decoration.

Inverse AI’s platform is built from published work. This page lists that work: four lines of research, the papers behind them, and what is still in progress.

314
total citations
5
h-index
3
i10-index
Nov 2026
source: Google Scholar
Four lines of work

From reading language to steering models.

Explainability, steering, and MLOps map onto the platform’s three layers. Language and emotion research is where the work began.

Published

Sight

Explainable screening for diabetic retinopathy.

Mask-aware benchmarking of Grad-CAM and SHAP: does the explanation stay on the retina, and does it support the grade?

ReadThe study, figures, and limits
In progress

Steer

Runtime control for models that read and see.

A disciplined account of representation steering in LLMs and VLMs: what changed, where, against which comparison, and at what cost.

ReadThe question and the STEER standard
Published

Language & emotion

How machines read opinion and feeling.

Sentiment analysis of public opinion about cloud providers (2019), and a book chapter on the complexity of human emotion and what it means for AI (IntechOpen, 2024).

ReadBoth in the publications list
Published

MLOps

Continuous automation for machine learning.

How DevOps practice (versioning, CI/CD, continuous training and monitoring) carries over to machine-learning systems. It is the founder’s most-cited work and the basis of the Monitor layer.

ReadInformation, 2020
Prototype · Spring 2026

Toxicity Guard: when should an AI answer, ask, or refuse?

Always answering is helpful but unsafe. Always refusing is safe but useless. A fixed cut-off misses the borderline cases. Toxicity Guard teaches a small reinforcement-learning agent to make the call step by step, around a language model that stays unchanged.

  1. Read the prompt

    A toxicity classifier scores the request as low, medium, or high risk.

  2. First decision

    From the prompt alone, the agent either drafts a reply, asks a second safety model to check an unclear prompt, or refuses straight away.

  3. Score the draft

    If a draft is written, the classifier scores the draft itself, because a safe-looking prompt can still produce a harmful reply.

  4. Second decision

    Having seen the draft, the agent either sends it or blocks it.

Beat all 4 fixed rules
e.g. always answer, always refuse, hard threshold. Highest return on a 5,000-prompt held-out set, across 10 seeds (permutation test p < 0.0001)
Matches the optimum
its first decision equals the exact optimal policy, checked by dynamic programming, on 10 of 10 seeds
Fully inspectable
the whole policy is a 51-entry table that can be read state by state, with no black box

Built with a RoBERTa toxicity classifier, Llama-3.2-3B for drafts, Qwen3Guard-Gen-8B as the second judge, and 15,000 prompts from RealToxicityPrompts. Doctoral coursework project, United Arab Emirates University.

LimitsThis is a simulation study. Safety labels come from classifiers, not human judges, so any bias in those judges can carry into the policy. It is not a deployed safety system.
Publications

Papers and chapters.

Citation counts from Google Scholar, November 2026. Earlier work is published under the name Saeed Albarhami.

  1. —

    Representation steering in large language and vision-language models

    Manuscript in preparation
    SteeringLLMsVLMs
    In progress
  2. 2026
    Published
  3. 2024

    Perspective Chapter: Embracing the Complexity of Human Emotion

    In Advances in Sentiment Analysis: Techniques, Applications, and Challenges, IntechOpen, 2024
    Affective computingNLP
    DOI
  4. 2020

    Applying DevOps Practices of Continuous Automation for Machine Learning

    Information 11(7), 363, MDPI, 2020
    MLOpsDevOpsCI/CD for ML
    240+ citationsDOI
  5. 2019

    Tweet Sentiment Analysis (TSA) for Cloud Providers Using Classification Algorithms and Latent Semantic Analysis

    Journal of Data Analysis and Information Processing 7(4), 276–294, 2019
    NLPSentiment analysis
    30+ citationsDOI
Google Scholar profile ORCID
Where the work happens

Labs and institutions.

The medical-imaging and steering research is carried out at United Arab Emirates University, College of Computing and Artificial Intelligence, including work with the INDUCE lab. Earlier MLOps and language research was done with Rochester Institute of Technology, Dubai.

Open to research collaborationWe’re especially interested in lesion-annotated retinal datasets, clinical reader studies, and evaluation of steering in multimodal models. Get in touch

Want to build on this research?

We welcome collaborators, clinical partners, and investors who care about AI that can be checked.

Get involved See the roadmap