What it does
FairLens lets users inspect datasets or model outputs for group-wise outcome differences.
Case Study
FairLens
AI-powered bias and fairness analysis tool
FairLens helps analyze datasets and model outputs for possible unfairness by comparing outcomes across sensitive attributes and generating AI-assisted explanations.
Overview
Making fairness analysis easier to inspect.
Why it was built
It was built as an MVP to make bias checks more approachable for small teams and learners.
Who it helps
It is useful for students, developers, and early-stage teams exploring fairness in ML tools.
Problem
Bias can be hard to see until outcomes are compared clearly.
ML systems can produce unfair outcomes when different groups receive different results.
Many beginners and small teams do not have an easy way to inspect bias in datasets or predictions.
FairLens was built as a simple MVP to make fairness analysis more understandable.
My Contribution
Full-stack work across analysis, validation, charts, and AI insights.
Key Features
Focused tools for beginner-friendly fairness inspection.
Upload / demo dataset analysis
Users can bring their own CSV or try the flow with a demo dataset.
Sensitive attribute selection
The UI guides users to choose the group attribute used for comparison.
Target and prediction handling
Target and optional prediction columns define the analysis mode.
Selection-rate metrics
The backend computes group-wise rates to make outcome differences visible.
Bias visualization charts
Charts turn the computed metrics into easier-to-scan comparisons.
AI-generated explanation
Gemini-assisted summaries explain what the results may suggest.
Beginner-friendly results
The output is written for clarity instead of assuming ML expertise.
Technical Approach
A clear flow from dataset choice to explainable results.
System Architecture
A high-level overview of how FairLens processes datasets and generates explainable fairness insights.
The architecture combines a Next.js frontend, FastAPI backend, fairness analysis engine, and Gemini-powered explanation pipeline.
Challenges & Fixes
Small product decisions made the MVP clearer and more reliable.
Challenge
Invalid column combinations
Fix
Added validation for target, sensitive, and prediction columns.
Challenge
Prediction column confusion
Fix
Made prediction optional and handled dataset vs model mode clearly.
Challenge
AI insight failures
Fix
Added fallback messages and improved timeout/model configuration.
Challenge
Chart edge cases
Fix
Handled all-zero and single-group scenarios.
Challenge
Deployment fetch issues
Fix
Documented frontend/backend integration limitations.
What I Learned
Designing AI tools for clarity, not just features.
- How fairness metrics can be made understandable through UI.
- Importance of validation before analysis.
- Difference between dataset-level and model-level bias analysis.
- Handling unreliable AI responses gracefully.
- Designing AI tools for clarity, not just features.
Future Improvements
Where FairLens could grow next.
- Expand advanced fairness metrics
- Make reports more actionable (not just downloadable)
- Add model-to-model comparison
- Improve real-time user guidance
- Enhance chart explanations & insights
Explore More