In-progress project: Building a neural network to learn an individual's style of speaking, and respond in their manner.

Current State: Trained on my iMessages from the past year (many of which are with my mom). ~60% accuracy in letter-by-letter message generation. Disclaimer: I am not responsible for what the neural network generates! Live Chat.

• Training messages obtained with SQL
• Recurrent Neural Network (RNN) built with TensorFlow
• Flask app responds with generated messages to a Twilio SMS webhook
• AWS EC2: Trained RNN on p2.xlarge instance, message generator deployed through t2.micro instance

Data was extracted from the iMessage sqlite3 database, located in /Library/Messages/chat.db.

Getting all messages from yourself:

SELECT text FROM message WHERE is_from_me = 1;

Getting all messages from another person:

-- get their handle_id
SELECT * FROM handle WHERE id="(their_phone_number)";
-- use it to extract their messages
SELECT text FROM message WHERE handle_id = "(handle_id)" AND is_from_me = 0;

Save the ascii messages to a .txt file and format it (ie; deal with emojis).

With dataset.py, process the text data with parameters from config.py into batches. The input shape for the recurrent neural network is: (None, 32, 62, 256).

Now draw the rest of the owl

When a text is sent to our Twilio webhook, it hits out app, which generates a message and responds.

Chatbot available through both SMS and online chat.

Our app is very basic. It has two backend routes, one to respond to texts via Twilio and one that responds to messages on the online app version.

@app.route("/sms", methods=['GET', 'POST'])
def sms_reply():
    resp = MessagingResponse()
    rnn_message = generate_message(459)
    resp.message(rnn_message)
    return str(resp)

@app.route("/message", methods=['GET'])
def message_reply():
    return str(generate_message(459))

The current step is to generate responses of varying natural lengths 🤔.

• Add a second layer to the model
• Improve RNN performance with LSTM cells
• In progress: Build web app version with React
• In progress: Generate responses of varying lengths
• Look into Natural Language Processing (NLP)
• Increase training data