Agente de Triagem de Suporte ao Cliente
Juadsuarezsan/support-triage-agent
Triagem inteligente de suporte: classifier + Claude + retrieval
Ver no GitHubAgent que automatiza a triagem de tickets de suporte combinando um classificador DistilBERT fine-tuned com LoRA para detecção de intenção, raciocínio Claude para análise de sentimento/prioridade, e busca semântica em Qdrant de tickets similares resolvidos. Decide automaticamente se resolve, sugere revisão humana ou escala o ticket.
Instalação
npx convez-ai Juadsuarezsan/support-triage-agent1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automatizar roteamento de tickets de suporte em massa
- Reduzir tempo de resposta identificando intenção e prioridade automaticamente
- Sugerir resoluções baseadas em tickets históricos similares
Veredito

Projeto ambicioso e bem estruturado que demonstra pipeline realista de LLM + ML fine-tuning para produção. Status ainda em planejamento (roadmap completo), mas arquitetura é sólida e aplicável a sistemas Zendesk/Intercom. Excelente para quem quer aprender orquestração LangGraph + retrieval + modelos classificadores.
README
A tradução para o português deste README ainda está na fila.
Project 02 — Customer Support Triage Agent
DistilBERT fine-tuned with LoRA for intent classification + Claude for reasoning + similar-ticket retrieval. Tickets arrive via email/Slack/chat, the system classifies, drafts a solution, and decides auto-resolve / suggest / escalate.
Industrial use case: Intercom, Zendesk, Freshdesk, HubSpot.
What this project does
Automatic triage of customer support tickets. Combines a fine-tuned BERT classifier (cheap, fast) with Claude reasoning (expensive, smart) plus retrieval of similar resolved tickets. Routes each ticket to auto-resolve, suggest-to-human, or escalate based on a confidence score.
Architecture
Ticket entra (email/chat/Slack)
│
▼
[Intent Classifier] DistilBERT + LoRA fine-tuned on Bitext (27 intents)
│ → intent + confidence
│
▼
[Sentiment + Priority] Claude structured output (Pydantic)
│ → sentiment, priority P0-P3, urgency_score
│
▼
[Similar Tickets] Qdrant semantic search
│ → top 5 resolved tickets with their resolutions
│
▼
[Solution Drafter] Claude
│ → proposed response grounded in past resolutions
│
▼
[Confidence Estimator]
│ score = f(intent_conf, sentiment, similarity, priority)
│
├─ > 0.85 → AUTO-RESOLVE
├─ 0.6-0.85 → SUGGEST (human reviews)
└─ < 0.6 → ESCALATE
│
▼
[Audit Logger] Postgres
Roadmap to v1.0.0
- Load Bitext dataset (27K queries × 27 intents), 80/10/10 stratified split
- LoRA fine-tune DistilBERT on Bitext train split, evaluate macro-F1
- Publish fine-tuned model to HuggingFace Hub with full model card
- Ingest historical resolved tickets into Qdrant (Twitter CS dataset for evaluation)
- LangGraph agent: classifier → sentiment → similar tickets → drafter → confidence scorer
- Eval set on Twitter dataset (200 end-to-end conversations)
- Compare: zero-shot Claude vs classifier-only vs hybrid pipeline
- Next.js Kanban UI showing tickets through states (pending / drafted / suggested / resolved / escalated)
- LangSmith public trace gallery (≥30 examples)
- Confusion matrix visualized in README
Stack
| Layer | Technology |
|---|---|
| Classifier | distilbert-base-uncased or roberta-base + PEFT/LoRA |
| Model hosting | HuggingFace Hub (with full model card) |
| LLM | Claude Sonnet 4.5 |
| Embeddings (for retrieval) | sentence-transformers/all-mpnet-base-v2 |
| Vector store | Qdrant |
| Orchestration | LangGraph |
| State | PostgreSQL |
| Frontend | Next.js 14 + shadcn/ui (Kanban) |
| Observability | LangSmith (public traces) |
Definition of Done — project-specific
- DistilBERT fine-tuned and published to HF Hub with model card (benchmarks, use cases, limitations, license)
- Comparative table: zero-shot Claude vs fine-tuned-only vs hybrid system, with macro-F1 / auto-resolve / latency / cost columns
- Confusion matrix in README
- Demo: Kanban with pre-loaded tickets, can drag through states
- LangSmith gallery linked from README
- Error analysis on the 10 worst-performing intents documented in
docs/error_analysis.md
Plus the 12 universal blocks of the DoD.
License
MIT.