Self-improving AI support triage. Resolving tickets, intelligently.
Live → loop-desk.vercel.app
LoopDesk is a production-grade multi-agent support triage system. It classifies, routes, and auto-resolves customer support tickets using LangGraph and Claude Sonnet 4.6. What makes it different — it gets smarter every time a human corrects it. No retraining. No fine-tuning.
Submit any support ticket. The agent classifies it, searches the relevant documentation, scores its own confidence, and either resolves it automatically or escalates to a human reviewer with a pre-written context summary.
Customer submits ticket
↓
LangGraph classifier (GPT-4o-mini)
↓ routes to
┌───────────────────────────────┐
│ Billing Agent │
│ Technical Agent │ ← Claude Sonnet 4.6 + ChromaDB RAG
│ General Agent │
└───────────────────────────────┘
↓
Confidence scoring (GPT-4o-mini)
↓
High confidence → Auto-resolved
Low confidence → Escalated + context summary
↓
Human reviews via Reviewer Dashboard
↓
Correction saved to Supabase
↓
Agent learns — few-shot injected on next similar ticket
Agent answers ticket
↓
Human marks it correct or wrong
↓
Correction stored in Supabase (permanent, survives deploys)
↓
Next similar ticket → correction injected as few-shot example
↓
Agent routes correctly without any code changes
This mirrors RLHF at a systems level — without touching model weights.
Tested across 21 ticket types including edge cases:
| Metric | Result |
|---|---|
| Core support tickets resolved correctly | 18/18 |
| Gibberish / off-topic escalated correctly | 5/5 |
| Corrections needed to fix routing errors | 8 |
| Retraining required | 0 |
| Hallucinations detected | 0 |
| Layer | Technology |
|---|---|
| Agent Orchestration | LangGraph |
| RAG + Prompting | LangChain |
| Specialist LLM | Claude Sonnet 4.6 (Anthropic) |
| Classifier LLM | GPT-4o-mini (OpenAI) |
| Vector Store | ChromaDB |
| Corrections Store | Supabase (PostgreSQL) |
| Observability | LangSmith |
| Workflow Automation | n8n |
| API | Flask + Flask-CORS |
| Frontend | HTML / CSS / JS |
| Backend Deployment | Render |
| Frontend Deployment | Vercel |
loopdesk/
├── webhook.py ← Flask API — main entry point
├── app.py ← Streamlit UI (backup demo)
├── index.html ← Landing page with live demo
├── static/
│ └── Logo.png
├── agent/
│ ├── graph.py ← LangGraph state machine
│ ├── classifier.py ← Classifier node (GPT-4o-mini)
│ ├── confidence.py ← Confidence scoring + escalation
│ └── agents/
│ ├── billing.py ← Billing specialist (Claude Sonnet 4.6)
│ ├── technical.py ← Technical specialist (Claude Sonnet 4.6)
│ └── general.py ← General specialist (Claude Sonnet 4.6)
├── rag/
│ ├── loader.py ← Document loader + chunker
│ └── retriever.py ← ChromaDB retriever
├── memory/
│ └── corrections.py ← Supabase-backed correction store
└── docs/
├── billing_faq.txt
├── technical_faq.txt
└── general_faq.txt
- Python 3.10+
- OpenAI API key
- Anthropic API key
- Supabase project
- LangSmith API key
git clone https://github.com/gitshiven/LoopDesk
cd LoopDesk
python -m venv venv
source venv/bin/activate
pip install -r requirements.txtOPENAI_API_KEY=your-openai-key
ANTHROPIC_API_KEY=your-anthropic-key
SUPABASE_URL=your-supabase-url
SUPABASE_ANON_KEY=your-supabase-anon-key
LANGSMITH_API_KEY=your-langsmith-key
LANGSMITH_TRACING=true
LANGSMITH_PROJECT=loopdesk
CREATE TABLE corrections (
id SERIAL PRIMARY KEY,
message TEXT NOT NULL,
wrong_category TEXT,
correct_category TEXT NOT NULL,
note TEXT,
created_at TIMESTAMP DEFAULT NOW()
);python webhook.pyBackend — Render: https://loopdesk-pl8q.onrender.com
Frontend — Vercel: https://loop-desk.vercel.app
Note: Render free tier spins down after inactivity. Use UptimeRobot to keep it awake.
Replace files in /docs with your own FAQ and policy documents, delete chroma_db/, rebuild the vector store, and redeploy. The agent learns your business in minutes.
Shiven Singh — github.com/gitshiven
MIT License