LoopDesk
Resolving tickets, intelligently.
Issue No. 001  /  2026

Close
The
Loop.

An AI triage agent that classifies, routes, and resolves support tickets — then gets smarter every time a human corrects it.

System / Active
agent.triage — live Running
LoopDesk v0.1 — RAG loaded (8 chunks)
Corrections store: active
 
customer "Charged twice this month"
classifier → BILLING (conf: 100%)
billing_agent searching docs... match found
confidence 1.0 > 0.75 → AUTO-RESOLVE
 
✓ resolved powered by Claude Sonnet 4.6
$
01
67→100%
Accuracy w/ Learning
2
Corrections to Fix Errors
Agent Types
0
Retraining Required
Classify Route Resolve Learn LangGraph RAG ChromaDB Claude Sonnet 4.6 Auto-Learning Human-in-the-Loop Classify Route Resolve Learn LangGraph RAG ChromaDB Claude Sonnet 4.6 Auto-Learning Human-in-the-Loop
§ 01 — How It Works

From message
to resolution

01 →
Ticket Arrives
Customer submits a support message in plain language. No categories, no forms. Just words.
Input Layer
02 →
Classify & Route
LangGraph classifier node reads intent, routes to the correct specialist agent — Billing, Technical, or General.
LangGraph Node
03 →
Resolve or Escalate
High confidence → answer sent automatically. Low confidence → human gets a pre-written handoff summary.
Output Layer
§ 02 — Auto-Learning

Gets smarter
every correction

1
Agent answers a ticket with a confidence score attached
2
Human reviewer marks it correct or wrong — adds a note if wrong
3
Correction stored and injected as a few-shot example on future similar tickets
4
Accuracy improves visibly over time — no retraining, no fine-tuning

Mirrors RLHF at a systems level — without touching model weights. Pure prompt engineering at production scale.

// Live Demo
Running on NovaPay — a fictional payments company with self-curated support docs. Real AI. Real responses.
Want this for your business? We replace NovaPay docs with your knowledge base. Contact: hello@loopdesk.ai
§ 03 — Live Demo

Try it
right now

Describe your issue
First request may take up to 60 seconds — server warming up.
Try an example
Double charge 502 error Export data Webhook issue Enable 2FA Cancel subscription
Review agent decisions. Correct mistakes to improve accuracy.
No tickets yet — submit one in the Customer tab
Agent Response
Waiting for a ticket...
Category
Confidence
Status
Response
§ 04 — Features

Built for production,
not demos

F-01
LangGraph Routing
Multi-node state machine with conditional routing. Each specialist has isolated RAG context and toolset.
F-02
RAG over Docs
Answers grounded in your docs via ChromaDB. No hallucinations — only what you define as ground truth.
F-03
Auto-Learning Loop
Every correction feeds back as a few-shot example. No retraining. Just smarter prompts.
F-04
Confidence Scoring
Knows when it doesn't know. Sub-threshold confidence triggers graceful escalation with a pre-written context summary for the human reviewer.
F-05 / F-06
LangSmith + Reviewer Dashboard
Full trace of every decision — chunks retrieved, routing reason, token cost. One-click reviewer dashboard. Deploy in minutes.
§ 05 — Stack

The stack

LangGraph
Agent Orchestration
LangChain
RAG + Prompting
ChromaDB
Vector Store
Claude 4.6
Specialist Agents
LangSmith
Observability
Render
API Deployment
"

The gap between a demo that impresses stakeholders and a system that handles real workload is exactly what LoopDesk is built to close.

LoopDesk — Self-improving AI triage, 2026

Ready to
close
the loop?

Open Source · MIT License

Built with LangChain, LangGraph, ChromaDB, and Claude Sonnet 4.6. Fully open source. Replace our docs with yours and deploy in minutes.