A professional deep learning web application that detects human emotions from text in real time using a Bidirectional LSTM (BiLSTM) neural network. Built with TensorFlow, Keras, Flask, and Bootstrap, this project transforms raw text into meaningful emotional insights.

Emotion-AI is a production-oriented sentiment understanding app that combines an NLP model with a Flask UI. In practice, the system works as a pipeline:

1. User enters free-form text in the web page (templates/index.html).
2. Backend sanitizes text (lowercasing + regex cleanup) before inference (app.py).
3. Inference strategy is selected at runtime:
   • Local model mode (TensorFlow assets loaded from disk) for full deep-learning predictions.
   • Remote inference mode (API forwarding) for lightweight cloud deployment such as Vercel limits.
   • Lexicon fallback mode for resilient predictions when ML dependencies are unavailable.
4. Emotion + confidence are returned as JSON and rendered instantly in the UI.

• Reliability: graceful fallback prevents total outage if model/runtime is unavailable.
• Deployability: remote inference keeps serverless package size small.
• Maintainability: model assets and app logic are separated (model_loader.py vs app.py).

• app.py → Flask routes, runtime mode switching, normalization, and fallback logic.
• model_loader.py → loading tokenizer/encoder/model artifacts.
• templates/ + static/ → web experience (UI, styles, interactions).
• requirements*.txt → dependency sets for local vs deployment scenarios.

• 🔮 Real-time emotion prediction

• 🧠 BiLSTM-powered deep learning model

• 🎭 Supports 7 emotions:
  😊 Joy | 😢 Sadness | 😡 Anger | 😨 Fear | 💙 Love | 😲 Surprise | 😐 Neutral

• 🌐 Interactive Flask web interface

1️⃣ Data Loading & Exploration
Dataset is ingested and split into 80% training / 20% testing.

2️⃣ Text Cleaning & Tokenization
Regex cleaning → Tokenization → Padding to fixed sequence length.

3️⃣ Model Architecture
Embedding → Bidirectional LSTM → Dropout → Softmax Output.

4️⃣ Training & Validation
Categorical Cross-Entropy + Adam Optimizer with Early Stopping & Checkpointing.

5️⃣ Evaluation & Optimization
Accuracy, Precision, Recall, Confusion Matrix & Hyperparameter tuning.

6️⃣ Deployment & Inference
Trained model (emotion_model.h5) integrated into Flask for real-time prediction.

• Input text via a modern UI
• One-click “Try a Sample” emotion buttons
• Live prediction display with confidence bar
• Fully responsive design

• Dataset loaded from .csv / .txt file
• Text tokenized using Keras Tokenizer
• Sequences padded for uniform input length

• Emotion labels converted to numeric form using LabelEncoder
• One-hot encoding applied for multi-class classification

Built using Keras Sequential API:

• 🔹 Embedding Layer – Converts words into dense vectors
• 🔹 Flatten Layer – Converts embeddings into 1D vector
• 🔹 Dense Layer – Learns complex patterns
• 🔹 Output Layer – Softmax activation for multi-class prediction

Compiled with:

• Optimizer: Adam
• Loss Function: categorical_crossentropy

• Dataset split using train_test_split
• Trained for 10 epochs
• Batch size: 32
• Validation data used to monitor performance

• Model predicts emotions for unseen text
• Example:

  Input: "I am feeling very nostalgic"
  Output: "Love"

flowchart LR
    subgraph DP[📂 Data Preparation]
        A["📦 Import Libraries"]
        B["📚 Load Dataset"]
        C["✂️ Preprocessing"]
    end

    subgraph LP[🏷️ Label Encoding]
        D["🔢 Encode Emotions"]
        E["📊 One-Hot Encoding"]
    end

    subgraph MT[🤖 Modeling & Training]
        F["🏗️ Build Model"]
        G["⚡ Train Model"]
    end

    subgraph PR[🔮 Prediction]
        H["📝 User Input"]
        I["🔍 Tokenize & Pad"]
        J["🎯 Predict Emotion"]
    end

    A --> B --> C --> D --> E --> F --> G --> H --> I --> J
    %% --- Styles ---
    style A fill:#FFD54F,stroke:#F57F17,stroke-width:2px,color:#000;
    style B fill:#4FC3F7,stroke:#0277BD,stroke-width:2px,color:#fff;
    style C fill:#AED581,stroke:#33691E,stroke-width:2px,color:#000;
    style D fill:#FFCC80,stroke:#EF6C00,stroke-width:2px,color:#000;
    style E fill:#FFE082,stroke:#F9A825,stroke-width:2px,color:#000;
    style F fill:#BA68C8,stroke:#4A148C,stroke-width:2px,color:#fff;
    style G fill:#FF8A65,stroke:#BF360C,stroke-width:2px,color:#fff;
    style H fill:#81D4FA,stroke:#01579B,stroke-width:2px,color:#000;
    style I fill:#B3E5FC,stroke:#0288D1,stroke-width:2px,color:#000;
    style J fill:#90CAF9,stroke:#0D47A1,stroke-width:2px,color:#000;

# Clone the repository
git clone https://github.com/shivareddy2002/emotion-ai.git
cd emotion-ai

# Install dependencies
pip install -r requirements.txt

# Run the app
python app.py

• Text Tokenization & Padding
• Multi-class Emotion Classification
• Deep Learning with Embeddings
• One-Hot Encoded Labels
• Validation-Based Training

• Python
• TensorFlow / Keras
• NumPy, Pandas
• Scikit-learn
• Jupyter Notebook

• 📱 Social Media Sentiment Analysis
• 🛒 Customer Feedback Classification
• 🛡️ Content Moderation
• 🧠 Mental Health Monitoring
• 💬 Chatbots & Virtual Assistants

This project demonstrates how Neural Networks can effectively understand and classify human emotions from text.
It highlights the power of NLP in real-world applications such as customer service, analytics, and well-being platforms.

Lomada Siva Gangi Reddy

• 🎓 B.Tech CSE (Data Science), RGMCET (2021–2025) | CGPA: 8.3
• 💡 Skills: Python, SQL, Snowflake, ETL, ML, DL, NLP, AI, Power BI
• 💼 SnowPro Core Certified | Data Engineering Intern (Boolean Data Pvt. Ltd.)
• 📍 Hyderabad, India | Open to Data & AI Opportunities

Contact Me:

• 📧 Email: lomadasivagangireddy3@gmail.com
• 📞 Phone: 9346493592
• 💼 LinkedIn 🌐 GitHub 🚀 Portfolio

Vercel serverless functions have strict bundle limits, and tensorflow-cpu is usually too large.

This repository now supports two modes:

1. Local mode: uses TensorFlow model files directly.
2. Vercel mode (default on Vercel): skips TensorFlow and forwards inference requests to a remote API.

1. Deploy a small Python API for model inference on a platform that supports heavy ML dependencies (Render/Railway/EC2).
2. In Vercel project settings, add:
   • USE_LOCAL_MODEL=false
   • REMOTE_INFERENCE_URL=https://your-ml-api.example.com/predict
3. Deploy this repo to Vercel. It installs only requirements-vercel.txt, which avoids large ML packages.

Use local dependencies when running model inference in the same process:

pip install -r requirements-local.txt
python app.py

If Vercel shows FUNCTION_INVOCATION_FAILED:

1. Open https://<your-domain>/health and verify:
   • use_local_model is false on Vercel
   • remote_inference_configured is true
2. In Vercel Project Settings → Environment Variables, set:
   • USE_LOCAL_MODEL=false
   • REMOTE_INFERENCE_URL=https://your-ml-api.example.com/predict
3. Redeploy after saving env vars.

If /predict cannot reach your remote inference API, the app now returns a lightweight fallback prediction so the UI still responds. For production-quality predictions, configure REMOTE_INFERENCE_URL to your ML service and keep USE_LOCAL_MODEL=false on Vercel.

Remote API payload should include either {emotion, confidence} or {label, score}. If payload is invalid, the app automatically falls back to lightweight local keyword prediction.

For highest quality predictions in production:

1. Prefer a dedicated remote inference service with your trained TensorFlow model.
2. Set REMOTE_INFERENCE_URL in Vercel to that endpoint.
3. Keep USE_LOCAL_MODEL=false on Vercel to avoid serverless crashes and size limits.
4. Use /health to verify runtime status (remote_inference_configured=true).

If remote service is unavailable, the app will safely fall back to the built-in lite model.

1. Ensure vercel.json includes templates/** and static/** in includeFiles for @vercel/python.
2. Keep USE_LOCAL_MODEL=false on Vercel unless TensorFlow and model files are actually available.
3. Check https://<your-domain>/health after deploy; it should return JSON (not crash page).
4. If root route fails, verify templates are deployed (missing templates cause serverless 500s).

A common root cause is different Vercel environment configuration between Preview and Production after merge:

• Preview deployment may have REMOTE_INFERENCE_URL configured while Production does not.
• Production may still have stale USE_LOCAL_MODEL=true from older deploys.
• Build packaging can differ if vercel.json changes were not merged cleanly.

After merging to main, verify Production env values in Vercel Project Settings and redeploy Production.