convez.ai
AgenteAutomaçãoAvançado
Juadsuarezsan

Agente de Triagem de Suporte ao Cliente

Juadsuarezsan/support-triage-agent

Triagem inteligente de suporte: classifier + Claude + retrieval

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent 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-agent

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

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.

Status LLM Fine-tune

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

  1. Load Bitext dataset (27K queries × 27 intents), 80/10/10 stratified split
  2. LoRA fine-tune DistilBERT on Bitext train split, evaluate macro-F1
  3. Publish fine-tuned model to HuggingFace Hub with full model card
  4. Ingest historical resolved tickets into Qdrant (Twitter CS dataset for evaluation)
  5. LangGraph agent: classifier → sentiment → similar tickets → drafter → confidence scorer
  6. Eval set on Twitter dataset (200 end-to-end conversations)
  7. Compare: zero-shot Claude vs classifier-only vs hybrid pipeline
  8. Next.js Kanban UI showing tickets through states (pending / drafted / suggested / resolved / escalated)
  9. LangSmith public trace gallery (≥30 examples)
  10. Confusion matrix visualized in README

Stack

LayerTechnology
Classifierdistilbert-base-uncased or roberta-base + PEFT/LoRA
Model hostingHuggingFace Hub (with full model card)
LLMClaude Sonnet 4.5
Embeddings (for retrieval)sentence-transformers/all-mpnet-base-v2
Vector storeQdrant
OrchestrationLangGraph
StatePostgreSQL
FrontendNext.js 14 + shadcn/ui (Kanban)
ObservabilityLangSmith (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.

Ver o original no GitHub

Bom para

Relacionados