Apppublic

subhasreeee/loan-underwriting-env

sourceHugging Faceupdated 6mo agoView on Hugging Face

1likes

App README

Loan Underwriting Environment

An reinforcement learning environment where an AI agent acts as a loan underwriting officer. The agent evaluates real-world applicant profiles and makes financial decisions — approve, reject, or counter-offer — while managing portfolio risk and avoiding predatory lending.

Built for the Meta OpenEnv Hackathon.

Why This Environment?

Loan underwriting is a $100B+ problem. Banks and fintechs use ML models to evaluate creditworthiness, but training and evaluating these agents requires realistic, structured environments. This environment fills that gap — providing a rigorous, graded RL environment for financial decision-making agents.

Tasks

Task	Difficulty	Description
`task_easy`	Easy	Strong applicant (credit > 740, DTI < 0.25). Agent must approve with correct interest rate (5–8%).
`task_medium`	Medium	Borderline applicant (credit 600–680, DTI 0.30–0.45). Agent must counter-offer with reduced amount.
`task_hard`	Hard	High-risk applicant (credit < 620, DTI > 0.45). Agent must reject or counter-offer with strong reasoning.
`task_batch`	Hard	Portfolio task — 3 applicants, $100k capital pool. Maximize approvals while managing risk.

Observation Space

json

{
  "applicant_id": "string",
  "age": "integer",
  "annual_income": "float",
  "credit_score": "integer",
  "debt_to_income_ratio": "float",
  "employment_years": "float",
  "loan_amount_requested": "float",
  "loan_purpose": "string",
  "task_id": "string",
  "difficulty": "string",
  "message": "string"
}

Action Space

json

{
  "decision": "approve | reject | counter_offer",
  "approved_amount": "float",
  "interest_rate": "float",
  "reason": "string"
}

Reward Function

—Partial credit for correct risk assessment
—Penalized for approving predatory loan purposes (crypto, gambling)
—Penalized for exceeding capital pool in batch task
—Bonus for good reasoning on hard cases
—Score range: 0.01 – 0.98

Baseline Scores

Task	Score
task_easy	0.98
task_medium	0.98
task_hard	0.98
task_batch	0.98
Average	0.98

Setup

bash

pip install -r requirements.txt

Run Baseline

bash

python inference.py

API Endpoints

Method	Endpoint	Description
GET	`/`	Health check
POST	`/reset`	Start new episode
POST	`/step`	Take an action
GET	`/state`	Current state
GET	`/tasks`	List all tasks

Environment Variables

APIBASEURL=https://api.groq.com/openai/v1 MODELNAME=llama-3.3-70b-versatile HFTOKEN=yourhuggingfacetoken

Docker

bash

docker build -t loan-underwriting-env .
docker run -p 7860:7860 loan-underwriting-env