ADI 5

We built a self-improving AI customer operations system that combines a live chatbot, voice-ready responses.

Comment

Inspiration

Customer support systems today are reactive, static, and require constant manual tuning. Businesses deploy chatbots and call-center automation, but these systems rarely improve on their own. We were inspired by the idea of building not just a chatbot — but a self-improving AI workforce that learns from its own conversations.

We asked: What if your customer support system had its own AI supervisor?

What it does

ADI 5 is a self-optimizing AI customer operations system.

It includes:

A live AI chatbot (WhatsApp-ready)

A voice-ready AI response system (via ElevenLabs)

A real-time logging and evaluation layer

A Supervisor AI that analyzes failed conversations and improves the system over time

The runtime agent handles customer conversations. Every interaction is logged and evaluated. A separate Supervisor Agent analyzes tone, intent, emotional signals, tool usage, and escalation patterns. It then generates structured improvements to the system prompt and behavior rules.

The result: The system continuously becomes smarter, more empathetic, and more effective.

How we built it

We built ADI 5 in layers:

n8n as the runtime orchestration layer for the customer-facing agent

Langfuse for observability, evaluation, and trace logging

Kiro (AWS) to develop the Supervisor Agent responsible for behavioral diagnostics and improvement logic

ElevenLabs for natural voice synthesis

Structured JSON-based communication between all services via HTTP APIs

The architecture follows a feedback loop:

Customer → Runtime Agent → Logging → Supervisor Agent → Prompt Improvements → Runtime Agent (v2)

This creates a controlled self-learning cycle.

Challenges we ran into

Designing a safe self-improvement loop without breaking production logic

Structuring Supervisor outputs so they are actionable and system-level, not just better replies

Balancing automation with human oversight

Avoiding overengineering while under time pressure

Integrating observability in a meaningful way instead of just logging data

The biggest challenge was making the system improve behavior — not just responses.

Accomplishments that we're proud of

Building a two-layer AI system: a runtime agent and a supervising agent

Implementing tone-aware analysis and empathy diagnostics

Designing structured prompt evolution instead of ad-hoc edits

Creating a clean feedback architecture using modern AI tooling

Demonstrating measurable behavioral improvement between versions

We’re especially proud that ADI 5 doesn’t just answer — it learns.

What we learned

Observability is critical before improvement.

Tone and emotional intelligence matter more than raw accuracy.

Self-learning systems must be structured and safe.

Iteration beats complexity.

Real innovation is in architecture and feedback loops, not just model choice.

We also learned how to move fast without losing system clarity under pressure.

What's next for ADI 5

Fully automated supervisor-to-runtime updates with approval workflows

Real-time tone adaptation during conversations

Multi-agent collaboration (Supervisor + QA + Strategy agents)

Deployment at scale on AWS infrastructure

Business-facing conversational analytics for owners

Continuous A/B testing of prompt strategies

Our long-term vision is to build AI customer systems that improve themselves responsibly and autonomously.

Built With

Share this project:

Updates