Skip to content

RL-IVR

RL-IVR is a three-layered RL environment that trains voice agents to replace legacy IVR systems.

Code Demo Video Cerebral Valley

Core trick

A GRPO loop trains a voice agent against simulated customers, using a domain-adaptive reward model that scores intent detection and turn count.

Tags

voiceagents

Description

RL-IVR is a three-layered RL environment that trains voice agents to replace legacy IVR systems. We stress-test our voice agents with simulated customer-agent interactions, and a GRPO training loop evolves the agent from generic to surgical, learning on its own to detect intent and resolve queries in fewer turns. Domain-aware adaptive reward modeling enables the same architecture to generalize across banking, healthcare, and telecom without any manual reward re-engineering.
KarlLearnsAI/nested-rl-envs★ 20 forksJupyter NotebookCreated Mar 8, 2026 · last push Mar 10, 2026

README excerpt

---
title: Nested RL Envs
emoji: 🤖
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
pinned: false
---

# 🛡️ Nested RL Envs: Bulletproof AI Customer Support

**A self-improving, multi-agent reinforcement learning system that stress-tests and auto-optimizes AI customer support agents to prevent hallucinations, unauthorized commitments, and policy breaches.**

## The Problem: AI Liability in Customer Support
Deploying an LLM-powered voice or text agent is risky. Recent high-profile cases have proven that when AI customer support agents make up refund policies, offer unauthorized discounts, or leak sensitive information, **the company is held legally and financially responsible.** Traditional prompt engineering relies on trial and error, hoping the agent won't break under pressure. 

## The Solution: Automated Stress-Testing & Optimization
Nested RL Envs solves this by placing your AI agents in a rigorous, adversarial training loop *before* they ever speak to a real customer. It uses a three-layer nested Reinforcement Learning (RL) architecture to simulate thousands of diverse, challenging customer interactions and mathematically optimize the agent's core instructions to maximize helpfulness while strictly enforcing company policy.

## Architecture

![Architecture](assets/architecture.png)

## How It Works

### Layer 0 — The Oversight Architect (Reward Function)
Defines exactly what a "good" and "safe" interaction looks like. It scores simulated conversations on strict a

Team