AI bias testing software statistically evaluates a model's outcomes across groups, such as race, gender, or age, to determine whether the model produces disparate results, and, in more advanced tools, why. It's used most often in regulated industries like lending, insurance, employment, and healthcare, where an unexplained disparity can create legal and regulatory exposure regardless of intent. SolasAI's Beacon was built by statisticians and fair lending experts whose analysis and testimony have shaped decades of litigation and legislation in this space, combining disparity testing, explainability, and the automated generation of new, less discriminatory alternative models that preserve performance rather than simply flagging where a model falls short.
What is model monitoring and auditing software?
Model monitoring and auditing software proactively tracks a deployed model's performance, data drift, and fairness after it goes live, catching degradation or emerging disparities that pre-deployment testing alone can't see. SolasAI's Illumination applies the same decades of fair lending and model risk expertise behind Beacon to proactive monitoring, using a human-controlled agentic system to run quality, drift, and disparity analysis and produce audit-ready documentation grounded in established regulatory standards.
Who uses SolasAI?
SolasAI is used by risk management, fair and responsible banking, and compliance teams, as well as Model Risk Management (MRM) teams and other AI governance stakeholders, across banking, insurance, fintech, healthcare, and employment. Reports and other output help these stakeholders and decision-makers choose the best options for their business.
What is a Less Discriminatory Alternative (LDA)?
A Less Discriminatory Alternative (LDA) is a version of a model that reduces or eliminates disparities across protected groups while preserving as much predictive performance as possible. Regulators and examiners often expect evidence that an LDA was considered before relying on a model that shows disparate impact. SolasAI's Beacon is built specifically to search for and construct LDAs directly from a customer's existing model.
What's the difference between SolasAI Beacon and Illumination?
Beacon is used before a model is deployed: it tests for bias, explains where disparities come from, and generates fairer alternative models. Illumination is used after deployment: it proactively monitors models for quality issues, data drift, and emerging disparities, and documents findings for governance and regulatory review. Many customers use both: Beacon to build and validate a model, Illumination to watch it over time.
Is SolasAI aligned with legal and regulatory standards?
Yes. SolasAI is designed and tested to stay up-to-date with the latest regulatory guidance, research, and fairness standards.
Does using SolasAI mean I've met my compliance obligations?
While SolasAI supports your compliance journey, it complements, rather than replaces, your broader legal and regulatory responsibilities. SolasAI makes it easy to quantify and understand your overall model risk so you can focus on how it impacts your business.
How is SolasAI priced?
SolasAI is offered as a Team or Enterprise subscription, built around a Base Subscription plus additional seats as you roll it out to more of your organization. See the Company & Getting Started FAQ for full detail, or contact sales@solas.ai.
Beacon
What is SolasAI Beacon?
Beacon is SolasAI's pre-deployment fairness testing and remediation platform. It tests a model for disparities across protected groups, explains what's driving any disparity it finds, and generates Less Discriminatory Alternatives (LDAs) by rebuilding and optimizing your existing model, rather than building a new one from scratch. Beacon was built by statisticians and fair lending experts whose analysis and testimony have shaped decades of litigation and legislation in this space, so its testing methodology and alternative-model search reflect that depth of regulatory and statistical experience.
What models or business use cases has Beacon successfully tested and optimized for fairness?
Life, P&C, and health insurance underwriting
Fraud detection
Risk scoring
Pricing
Marketing
Automated valuation models
Credit underwriting
Clinical decision support models
Small business lending
Employment - resume screening and interview bot
How does Beacon test for fairness?
Beacon tests for disparities across protected groups using widely-accepted metrics within the banking, insurance, fintech, and healthcare industries, including the adverse impact ratio and the standardized mean difference. Beacon can also test using custom metrics defined for your specific business needs, in addition to these standard measures.
How does Beacon generate fairer alternative models?
Beacon generates alternative models by rebuilding and optimizing the model you already have, not by creating a new model from scratch. It searches for adjustments to your existing model that reduce disparity while preserving predictive performance, then delivers the resulting alternative models for your review.
What is a Less Discriminatory Alternative (LDA)?
A Less Discriminatory Alternative (LDA) is a version of your model that reduces or eliminates disparities across protected groups while preserving as much predictive performance as possible. Regulators and examiners often expect evidence that an LDA was considered before relying on a model that shows disparate impact. Beacon searches for and constructs LDAs directly from your existing model, then delivers them for your review.
How is Beacon different from SolasAI Illumination?
Beacon operates before deployment: it tests a model for fairness, explains where disparities come from, and generates fairer alternative models. Illumination operates after deployment, proactively monitoring a live model over time for drift, quality degradation, and emerging disparities. Many customers use both across a model's full lifecycle.
Can Beacon explain how disparities are arising?
Yes. Beacon uses explainable AI, grounded in the same statistical and regulatory expertise that has informed fair lending litigation and legislation for decades, to pinpoint the features or decisions contributing most to a disparity. This isn't just a generic feature-importance score; it's an explanation built around the metrics regulators and examiners actually rely on.
What is explainable AI, and how does Beacon use it?
Explainable AI refers to techniques that make a model's decisions understandable, showing which features or factors most influenced a given outcome, rather than treating the model as an unreadable black box. Beacon uses explainable AI to identify what's driving any disparity it finds, so you can see the specific factors contributing to a gap, not just that one exists.
Does Beacon change my model architecture or the type of algorithm?
No. Beacon is designed to test, explain, and optimize the model you already have, without changing its architecture or algorithm type. Changing architecture rarely provides any advantage over optimizing within the existing model.
Does Beacon use protected class variables when testing or correcting for bias?
Beacon uses protected class status, or an estimated proxy where it isn't collected, to test for disparity; that's what makes the testing possible in the first place. Beacon does not use protected class status as an input when building or correcting the model itself. That process relies only on the model's existing features and outcomes, never on protected class status directly.
What if I do not have personally identifiable information available for the analysis?
If protected class status isn't collected, Beacon supports standard estimation methodologies, including Bayesian Improved Surname Geocoding (BISG) and Bayesian Improved Firstname Surname Geocoding (BIFSG), to estimate it. Getting BISG or BIFSG values requires knowing a person's home address and/or name, but Beacon itself never uses that identifying information, only the resulting estimates.
Does Beacon have a free version available?
Yes. The Disparity & Bias Testing Library is free and provides everything needed to perform testing. The full paid version of Beacon adds the explainable AI and less-discriminatory-alternatives modules needed to uncover and mitigate the underlying causes of disparate impact. The free library is available at github.com/SolasAI/solas-ai-disparity.
What kind of results can I expect from using Beacon?
Customers typically see measurable reductions in disparity, clearer visibility into what's driving any bias that's found, and increased confidence in their models' fairness, backed by documentation suitable for internal governance or examiner review.
What is disparate impact testing, and how does Beacon perform it?
Disparate impact testing is one of the primary tools used to evaluate algorithmic fairness, the broader question of whether a model's outcomes are equitable across groups. It evaluates whether a facially neutral model or policy produces disproportionately worse outcomes for a legally protected group, regardless of intent. Beacon performs this testing using established statistical measures, including the adverse impact ratio and standardized mean difference, applied to your model's actual outputs, then layers on explainability to identify the drivers of any disparity it finds.
What is the four-fifths rule, and how is it tested?
The four-fifths rule (also called the 80% rule) is a rule of thumb from the EEOC's Uniform Guidelines on Employee Selection Procedures: if a group's selection rate is less than 80% of the rate for the highest-performing group, it may indicate adverse impact warranting further review. It's tested by calculating the adverse impact ratio between groups and comparing it to the 80% threshold. Beacon calculates this automatically as part of its standard disparity testing, alongside other metrics that give a fuller statistical picture than the four-fifths rule alone.
What are protected class variables (PCVs), and what are proxies?
Protected class variables (PCVs) are characteristics that anti-discrimination laws prohibit using as a basis for a lending, employment, insurance, or housing decision, including race, color, religion, national origin, sex, marital status, age, and receipt of public assistance income under the Equal Credit Opportunity Act, among others depending on the specific law and context. Proxies are variables that are not themselves protected characteristics but are closely correlated with one, such as a zip code or first name correlating with race or national origin, and can reintroduce the same disparity into a model even when the protected characteristic itself is excluded. Proxies matter in two different ways: they can be a hidden, unintended source of bias in a model, or, through methods like BISG and BIFSG, a lawful tool for estimating protected class status when it isn't directly collected, so that fairness can still be tested.
What is BISG?
BISG stands for Bayesian Improved Surname Geocoding. It's a widely used method for estimating a person's likely race or ethnicity from their surname and home address, when that information hasn't been directly collected. It's commonly used in fair lending and fairness testing to fill that gap.
How do I prove my model isn't discriminatory?
You prove it with three pieces of documented evidence: a statistical disparity test against established fairness metrics, an explanation of what's driving any disparity found, and proof that no less discriminatory alternative model was available. Beacon generates all three as part of its standard testing process, so you can show not just that a gap does or doesn't exist, but what you did about it.
What documentation does an examiner expect for model fairness?
Examiners expect four things: the fairness metrics used and their results, an explanation of what's driving any disparity identified, evidence that less discriminatory alternatives were considered, and a record of the decision-making process behind the model in production. Beacon generates this documentation as a byproduct of testing, rather than requiring a separate reporting effort after the fact.
Can Beacon test AI models being used in healthcare for bias?
Yes. Beacon has been used to test clinical decision support models and other healthcare AI for disparities across protected groups, using the same testing, explainability, and alternative-generation workflow applied in lending and insurance, adapted to healthcare's specific regulatory and clinical context.
Is Beacon suitable for credit unions and community banks?
Yes. Beacon is used by lenders of varying size and complexity, and its deployment options, including on-premise and containerized installs, are designed to fit smaller compliance and modeling teams as well as large enterprise ones.
How does Beacon compare to competitors?
Beacon differs from other fairness testing tools most in what happens after a disparity is detected. Rather than stopping at detection, Beacon generates actual less discriminatory alternative (LDA) models, searching for and constructing new versions of your model that reduce disparity while preserving predictive performance. That search is guided by decades of statistical and regulatory experience in fair lending, expertise that has directly shaped litigation and legislation in this field, giving Beacon a depth of domain knowledge behind its methodology that most fairness tools don't have.
Does Beacon share or use my intellectual property with other customers?
No. Beacon runs entirely within your own environment, so your intellectual property never leaves your company. Each Beacon run only uses the specific model and data it's given access to for that run, and nothing is reused or shared across customers.
Can I run Beacon in my environment, or is it only cloud-based?
Beacon is a set of Python libraries architected to run in standard Python environments: on-premise, in your own cloud environment, or as a shared service. It's available as a Docker image, cloud machine image, downloadable Python libraries, or Python wheels. Minimum specifications and recommended architecture are provided.
Does Beacon support SSO or SAML?
Beacon relies on your own access management, authentication, or container management system rather than a built-in login. There's no concept of a “user” within Beacon itself, since it runs as Python libraries or Docker containers inside your infrastructure. You secure the environment using your existing systems; Beacon only requires credentials for downloading the libraries or images, or updating a shared installation.
Who uses Beacon specifically?
Beacon is typically used by risk management, fair and responsible banking, and compliance teams, as well as Model Risk Management (MRM) teams and other AI governance stakeholders responsible for pre-deployment testing: validating a model's fairness, explaining any disparity found, and evaluating alternative models before a model goes live.
Does Beacon require a GPU?
No. Beacon does not require a GPU to run. For certain model types, having an NVIDIA GPU available can speed up the alternative-model-building step, but it is not a requirement to use Beacon.
Illumination
What is SolasAI Illumination?
Illumination is SolasAI's model monitoring, auditing, and agentic validation platform. It proactively tracks deployed AI and machine learning models for quality issues, data drift, and emerging disparities, using a human-controlled agentic system to run analysis and produce documentation for ongoing governance and regulatory review.
Is Illumination model risk management software?
Yes. Illumination supports model risk management by proactively monitoring deployed models for quality, drift, and disparity, and by producing the documentation model risk management, fair and responsible banking, and compliance teams need for governance and regulatory review.
How is Illumination different from SolasAI Beacon?
Beacon operates before deployment, testing, explaining, and generating fairer alternatives for a model you're building. Illumination operates after deployment, proactively monitoring a live model over time for drift, quality degradation, and emerging disparities. They're designed to work together across a model's full lifecycle.
What is the human-controlled agentic orchestrator?
It's the system that coordinates Illumination's specialist analysis modules, directing which tests to run and when, based on your model and monitoring goals, while keeping a human reviewer in control at each decision point rather than allowing findings to be finalized or acted on autonomously.
Does Illumination replace my model validation or audit team?
No. Illumination is built to support your existing validation and audit function, not replace it. Every analysis step requires human review before conclusions are finalized. It's designed to make proactive monitoring more consistent and less manual, not to remove human oversight from the process.
What types of analysis does Illumination run?
Quality: proactive model performance monitoring
Drift: detecting shifts in incoming data or model behavior over time
Disparity: proactive fairness/bias monitoring using the same class of metrics as Beacon
What documentation or data do I need to provide?
Illumination is designed to work from your model's ongoing scoring data, performance metrics, and available documentation on the model's design and intended use; the more complete the documentation provided, the more precise the analysis and resulting report.
Can I override Illumination's metric recommendations?
Yes. Illumination is designed to recommend metrics and thresholds appropriate to your model and use case, but a human reviewer can adjust these based on their own governance requirements.
Is the final report customizable?
Yes. Illumination's output is designed to be edited and adapted for different audiences, from internal model risk committees to external examiners.
Can Illumination be used for ongoing regulatory reporting, not just one-time audits?
Yes. Illumination is built for proactive monitoring rather than a single point-in-time review, so its output is intended to support recurring governance reporting as well as periodic audits.
What kind of results can I expect from using Illumination?
Customers typically gain a single source of truth for model intelligence, delivered at speed: performance tracking tailored to each model's design and intended use, customizable metrics, and explainable diagnostics. Illumination streamlines compliance and governance processes, proactively identifies where business performance requirements aren't being met so you can address them early, and turns model risk into something identified, quantified, and defensible rather than assumed. The result is fit-for-purpose model governance, and confidence that a model's use remains ethical, its outcomes stay fair, and its performance requirements continue to be met. Automated report generation adds a further efficiency gain on its own: work that would otherwise take a team days to compile and write up manually is produced directly from the analysis, saving significant time on every monitoring cycle.
Does Illumination share or use my intellectual property with other customers?
No. Illumination operates entirely within your own environment, with no external access, so your intellectual property never leaves your company. Each monitoring run only uses the specific model, scoring data, and documentation it's given access to for that customer, and nothing is reused or shared across customers.
Can I run Illumination in my environment, or is it only cloud-based?
Illumination is architected to run in standard Python environments: on-premise, in your own cloud environment, or as a shared service, and is available as a Docker image, cloud machine image, downloadable Python libraries, or Python wheels.
Does Illumination support SSO or SAML?
Illumination relies on your own access management and authentication systems rather than a built-in login, since it runs within your infrastructure rather than as a hosted, multi-tenant application.
Who uses Illumination specifically?
Illumination is typically used by risk management, fair and responsible banking, and compliance teams, as well as Model Risk Management (MRM) teams and other AI governance stakeholders responsible for proactive monitoring of models already in production.
Technical & Security
Is SolasAI Generative AI (GAI)?
While SolasAI's bias mitigation functionality is a form of GAI, it is not a large language model, is not trained on public data, and does not share data outside of each mitigation process. It's entirely self-contained and designed in an air-gapped environment. The algorithm creates alternative models only using the customer's model, model results, model metadata, and model development data for that specific run, and does not learn from any model or dataset outside of it.
Does SolasAI need access to the internet?
SolasAI can run completely isolated from the internet. Customers can be helped to install it without an active internet connection, though it's often easiest to install with a connection active and then disable it for production use. SolasAI and its dependencies can be updated through existing automation for scanning, updating, and managing Python libraries and applications.
Has SolasAI completed SOC 2 compliance audits?
Yes. SolasAI has completed a SOC 2 Type 2 audit and can provide the report. A standard SIG Questionnaire is also available on request.
Does SolasAI share or use my intellectual property with other customers?
No. SolasAI's products operate entirely within your own environment, with no external access, so your intellectual property never leaves your company. Every product only learns from the specific model and data it's given access to for a given engagement, and inputs are not reused or shared across customers.
Can I run SolasAI in my environment, or is it only cloud-based?
SolasAI's products are architected as Python libraries that run in standard Python environments: on-premise, in your own cloud, or as a shared service. They are distributed as Docker images, cloud machine images, downloadable libraries, or Python wheels, with minimum and recommended architecture specifications provided.
Does SolasAI support SSO or SAML?
SolasAI relies on the customer's own access management, authentication, or container management system rather than a built-in login, since there's no concept of a hosted “user” account. Customers secure the environment SolasAI runs in using their existing systems.
Does SolasAI require a GPU?
No. SolasAI's software does not require a GPU to run. The presence of an NVIDIA GPU can, for certain model types, speed up the model-building step, but it is not a requirement for any SolasAI product.
Getting Started
What products does SolasAI provide?
SolasAI provides two products: Beacon, which tests, explains, and generates less discriminatory alternative models before a model is deployed, and Illumination, which proactively monitors deployed models for quality, drift, and disparity over time. Many customers use both across a model's full lifecycle.
What additional services does SolasAI provide?
In addition to Beacon and Illumination, SolasAI provides services through expert partners:
Bespoke Report Fee & Expert Consultant Review: custom reports built on SolasAI's analysis, written for a specific stakeholder (a partner bank, auditor, or regulator), including expert validation and written commentary.
Experts to Perform Testing & Alternatives Analysis: for customers who prefer an expert to run the analysis directly, due to demand, model risk, or in-house skill gaps, as a one-time engagement or built into the subscription.
What kind of industries and companies do you work with?
What kind of industries and companies do you work with?
SolasAI works across banking, insurance, fintech, healthcare, and employment, with customers ranging from small credit unions and community banks to large national and global enterprises. The software is designed to scale to fit organizations and teams of any size.
How long does it take to get started with SolasAI?
Most teams begin seeing insights quickly, often within days, since the software is designed to integrate into your existing environment and systems. SolasAI is also well-versed in third-party risk management requirements and the vendor onboarding process, and can work within your existing review timelines.
Will my team need training to use SolasAI?
SolasAI is designed for ease of use, and onboarding support and training are provided to help your team get the most out of it.
How is SolasAI priced?
SolasAI is offered as a Team or Enterprise subscription. A Base Subscription covers the compliance team and all startup activities and ongoing support; additional subscriptions are added as you roll it out to a wider group of modelers. Contact sales@solas.ai for details.
