Python Developer InternCodec Technologies Pvt. Ltd. · 2026
SentiVision-AI — Multimodal Sentiment & Emotion Platform
A multimodal AI application combining transformer-based text sentiment with real-time facial emotion recognition.
3 Models
Sentiment Ensembles
RoBERTa, VADER, TextBlob
7 Classes
Facial Emotions
Joy, Sadness, Anger, Surprise, etc.
Multimodal
Input Modalities
Text, Reviews, Video Feed
System Architecture & Dataflow
Text/Webcam -> NLP Sentiment Pipeline & Computer Vision Emotion Detector -> Synthesis Engine -> Chart.js Dashboard
Pipeline Layers //
- ✦Text analysis: Twitter-RoBERTa transformer ensemble paired with VADER and TextBlob lexicons
- ✦Vision analysis: Haar-cascade face localization and emotion classification across live webcam frames
- ✦Review extraction: Automated text parsing from discussion threads and batch CSV files
- ✦Data persistence: Session-level MongoDB document storage and PDF analytics export
Applied Technologies & Libraries //
PythonFlaskTransformersRoBERTaOpenCVVADERTextBlobMongoDB
# SentiVision-AI: Multimodal Emotional Intelligence
Developed as my second major deliverable during the Python Developer Internship at **Codec Technologies Pvt. Ltd.**, SentiVision-AI merges two distinct modalities of human emotional expression into a unified computational dashboard.
## The Problem
User sentiment is rarely expressed through text alone or facial cues alone. Analyzing customer feedback, product reviews, or user interviews requires inspecting both the lexical sentiment of what was written/said and the emotional reactions visible in facial expressions.
## Engineering Highlights
- **Transformer NLP Ensemble**: Integrated a fine-tuned `cardiffnlp/twitter-roberta-base-sentiment-latest` transformer model for high-accuracy nuance, supported by VADER for fast rule-based sentiment scoring.
- **Real-Time Facial Emotion Detection**: Built an OpenCV Haar-cascade face detection pipeline feeding into a deep convolutional emotion classifier evaluating 7 micro-expressions in real time.
- **Multimodal Synthesis**: Reconciled divergent sentiment signals into an overall confidence index, charting trends over multi-minute sessions.