Carely

Emotion-aware AI assistant for customer-support conversations

Carely는 고객 상담 내용을 분석해 감정과 의도를 파악하고, 상담사가 사용할 응답 가이드와 안정 피드백을 생성하는 AI 상담 지원 프로젝트입니다.

This repository is a documentation-first portfolio of my contribution to the AI-BE team repository. The original team implementation and its commit history remain available at the link above.

Problem and Approach

Customer-support conversations can change tone quickly. A useful assistant must track emotion over time, understand the customer's intent, and generate a response grounded in service policies instead of reacting to a single sentence.

flowchart LR
    A[Customer utterance] --> B[KoBERT emotion classifier]
    A --> C[Intent classifier]
    B --> D[Session-level emotion smoothing]
    C --> E[Response guide agent]
    D --> E
    F[Policy knowledge] --> E
    D --> G[Agent-calming feedback]
    E --> H[FastAPI response]
    G --> H

Loading

My Contribution

• Integrated the audio-processing and Korean emotion-classification agent flow during development, including STT-provider iterations.
• Added the KoBERT emotion model, model-loading path, and probability output used by downstream agents.
• Implemented session-level emotion smoothing and confidence-based neutral handling to reduce unstable responses to short utterances.
• Connected the emotion result to the FastAPI request/response pipeline and agent orchestration.
• Resolved file-upload, schema, dependency, model-loading, and Render deployment issues across the team backend.

These contributions are traceable in 25 public commits authored under @hozziii.

Key Features

• Korean emotion classification with a fine-tuned KoBERT model
• Confidence-aware neutral-emotion handling
• Moving-average emotion smoothing maintained per conversation session
• Intent classification and policy-aware response generation
• Separate customer-response and counselor-calming messages
• FastAPI endpoint with typed Pydantic responses

Tech Stack

Area	Tools
Language and API	Python, FastAPI, Pydantic
Machine learning	PyTorch, Hugging Face Transformers, KoBERT
Agent pipeline	LangChain, OpenAI-compatible LLM APIs, FAISS
Audio experiments	STT-provider integration during development
Deployment	Uvicorn, Render

Repository Scope

This portfolio repository intentionally contains project documentation rather than a copied team codebase. Model files, service credentials, and deployment configuration are not duplicated here. For implementation details, see the original AI backend.

What I Learned

• Designing stable emotion signals for multi-turn interactions
• Integrating an ML model into a typed web API and multi-agent workflow
• Debugging model artifacts, runtime dependencies, and cloud deployment paths
• Coordinating feature work through branches, pull requests, and shared code