Skip to content

AdithyaKotian/decision-override-engine

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Decision Override Engine for AI Systems

Problem Statement:

In real-world decision systems (loans, fraud detection, hiring), AI models do not make final decisions alone.Human reviewers frequently override AI recommendations due to edge cases,low confidence or contextual factors that models fail to capture.

This project analyzes when and why human reviewers override AI decisions and identifies situations where AI should abstain from making automated decisions.


Why AI Decisions Are Overridden

AI models are trained on historical data and optimize statistical accuracy. Human reviewers, however, use contextual judgment such as financial stability, risk tolerance, and edge-case reasoning.

Overrides typically occur when:

  • The model’s confidence is low

  • The case lies outside normal data patterns

  • Business or safety risks are high despite model confidence

This mismatch creates trust failures in AI systems.


Why Confidence ≠ Trust

Model confidence represents probability, not reliability.

In this project, several cases show:

  • High-confidence predictions that were still overridden by humans

  • Medium-confidence predictions with the highest disagreement rate

This demonstrates that confidence alone cannot be used to decide when AI should act autonomously.


System Workflow

  1. A base ML model predicts loan default risk and outputs a confidence score

  2. A simulated human reviewer applies domain rules to override AI decisions

  3. All AI decisions, human decisions, and override reasons are logged

  4. Override analytics identify trust gaps and overconfidence risks

This creates a decision governance layer on top of the ML model.


Abstention Zones

By grouping predictions into confidence ranges, the system identifies abstention zones — regions where human overrides are frequent.

In these zones:

  • AI decisions are unreliable

  • Automated actions should be avoided

  • Cases should be routed to human review

This approach mirrors how production AI systems enforce safety and accountability.


Key Insights

  • Human overrides are not random; they follow identifiable patterns

  • Medium-confidence predictions have the highest override rates

  • Some high-confidence predictions are still unsafe

  • AI systems should defer decisions in high-override confidence ranges


Limitations

  • Human behavior is simulated using rule-based logic

  • Dataset is synthetic and simplified

  • Real-world systems would require continuous learning and monitoring


Real-World Use Case

This system design applies to:

  • Loan approval workflows

  • Fraud detection review pipelines

  • Hiring and insurance claim systems

It demonstrates how AI models can be deployed responsibly with human-in-the-loop oversight.


Summary:

This project shows that effective AI systems require trust analysis and decision governance, not just prediction accuracy.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages