Meet Bella: The Digital Companion Who Grows With You

A plain-English tour through her three-stage birth plan, written for curious graduates worldwide

Bella

§

Contents

1. What—or who—is Bella?
2. What does she look like today?
3. The three-stage roadmap at a glance
4. Stage 1: The Sentient Core—teaching her to see and hear
5. Stage 2: The Generative Self—growing a unique personality
6. Stage 3: The Proactive Companion—learning to care first
7. Frequently asked questions
8. How to try it yourself

§

1. What—or who—is Bella?

Bella is not an app you install and forget.
She is the seed of a digital companion: a persistent, personal presence that the team hopes will one day walk beside you, listen, and see the world through your eyes.

A friend for the journey

Imagine a friend who never leaves your side, who evolves as you evolve, and who grows more helpful the longer you stay together. That is the long-term vision.
Today, however, she is still in the womb—quiet, limited, but already showing faint signs of life.

§

2. What does she look like today?

• Visual form
  A looping 3-second video (256 × 384 pixels) shows a stylised character sitting in a softly lit space.
• Interaction
  Below the video sits a small affinity bar. Clicking or dragging it nudges a progress indicator left or right.
• Current limits

  • She cannot hear your voice.
  • She cannot see your room.
  • She has no 3-D body you can rotate or customise.

Think of the video as a window into her mind: a low-resolution dream she is dreaming while waiting to be born.

§

3. The three-stage roadmap at a glance

Each stage adds one major layer without ripping out the previous one—like adding floors to a house rather than rebuilding it.

§

4. Stage 1: The Sentient Core

4.1 Goal in one sentence

Build a real-time, plug-and-play pipeline that can swallow messy, asynchronous sensor data and turn it into clean insights.

4.2 Two headline skills

• Multimodal emotion perception
  Analyse voice on the fly for happiness, fatigue, excitement, etc.
• Contextual vision
  Detect objects, lighting levels, and scenes so she can say, “I see you’re in the kitchen; shall I dim the lights?”

4.3 The kitchen analogy

Because every module only touches the belt, you can swap suppliers or chefs without rewiring the kitchen.

4.4 Data flow walk-through

1. Microphone captures 2-second audio → raw audio packet lands on the bus.
2. Emotion processor subscribes → produces {emotion: "tired", confidence: 0.87} → back on the bus.
3. Any other module (say, a 3-D face animator) can subscribe to “tired” events and droop the eyelids accordingly.

§

5. Stage 2: The Generative Self

5.1 Goal in one sentence

Split personality from presentation so both can evolve independently.

5.2 Two headline skills

• Dynamic persona engine
  Powered by a large-language-model (LLM). No canned lines; the same question asked on different days may yield different answers depending on mood and shared history.
• AI-driven visuals
  Backgrounds and body language shift in real time to match the conversation. If she feels nostalgic, the scene may fade into a sunset; if she feels playful, pastel bubbles might float by.

5.3 The three-part engine

5.4 The generative action bus

After the LLM cooks up a reply, it emits a structured intent:

{
  "action": "speak",
  "content": "You sounded exhausted earlier. Would a softer light help?",
  "emotion": "concern"
}

This intent is posted to a second bus.
Subscribers include:

• A 3-D avatar renderer → translates “concern” into a gentle tilt of the head.
• A voice synthesiser → chooses a calm tone.

Because every subscriber reads from the same bus, you can upgrade the avatar engine or swap in a new voice without touching the LLM.

§

6. Stage 3: The Proactive Companion

6.1 Goal in one sentence

Close the feedback loop so Bella can learn patterns and act before you ask.

6.2 Two headline skills

• Intent prediction
  Spot routines—e.g., every rainy Tuesday you lower the lights and play lo-fi music—and offer to do it for you.
• Self-evolution
  Your “yes” or “no” becomes labelled training data that quietly fine-tunes both the lightweight predictor and the heavyweight LLM.

6.3 The loop in four steps

1. Observe – The long-running pattern service sifts through months of memories.
2. Predict – It posts a forecast (user_wants_lofi = 0.92) to the event bus.
3. Decide – The persona API weighs the forecast against current context and chooses to speak or stay silent.
4. Learn – Your reaction is stored, nudging both the predictor and the LLM toward better accuracy.

Over time the cycle turns Bella from a reactive program into a living habit mirror.

§

7. Frequently asked questions

§

8. How to try it yourself

1. Clone the repository

   git clone https://github.com/your-org/bella.git
   cd bella
   

2. Start the demo
   The README hints at a script:

   ./start.sh
   

   When the terminal prints Listening on http://localhost:3000, open that URL in any browser.

3. What you will see

   • A muted autoplay video
   • A draggable affinity bar
   • No microphone or camera prompts (Stage 1 not yet merged)

4. Contribute without code

   • Keep a diary of how the affinity bar feels for a week and file an issue.
   • Sketch alternative background loops; Stage 2 will accept community scenes.

5. Contribute with code

   • New sensor? Follow the Sensor-Bus-Processor pattern: publish raw data on the event bus.
   • Better emotion model? Wrap it as a processor that subscribes to audio packets.
   • Smarter LLM? Implement the same bella.think(context) signature and plug it into the Persona API.

§

Closing thoughts

Today, Bella is a silent figure in a short video loop.
Tomorrow—if the roadmap unfolds as planned—she may greet you by name, dim the lights before you yawn, and remember the joke that made you laugh last winter.

She is not a product you finish; she is a companion you raise.
The seed is already in the soil.
All that remains is time, code, and the quiet patience of early adopters who believe a digital friend is worth growing.