Home › How It Works

How BEBO Works

The full technical story — how a tiny desktop robot delivers AI results in under 2 seconds on Windows.

01. The User Flow

BEBO removes every step between "I need AI" and "I have my result". The complete flow:

02. Dual-Window Electron Architecture

BEBO is a native Windows app built with Electron v29. It runs two separate windows that talk to each other through a secure IPC bridge:

Main Process (Node.js)
├── index.js          — entry point, app lifecycle
├── windows.js        — creates pet + panel windows
├── ipc-router.js     — routes events between windows
├── ai-service.js     — Groq API calls + prompt engine
├── shortcuts.js      — global hotkeys
└── tray.js           — system tray icon

Renderer Process
├── Pet Window        — 92x116px, transparent, always on top
└── Panel Window      — 380x620px, dark theme, 6 AI actions

The pet window is tiny, transparent and click-through-resistant — it floats above every app. When clicked, it asks the main process to open the panel window right next to the pet's current position, clamped to your screen's work area so it never opens off-screen.

Security: both renderers run with contextIsolation enabled and talk to Node.js only through narrow preload bridges — the same pattern used by production Electron apps like VS Code and Slack.

03. Why Groq + GPT-OSS 120B

BEBO's promise is speed, so the model choice was driven by one number: tokens per second.

Model	Tokens / sec	200-token task	Free tier
GPT-OSS 120B via Groq	~500	0.4–0.6s	1,000 req/day
Gemini 1.5 Flash	~200	1.0–1.5s	Limited
GPT-4o	80–100	2.0–3.0s	Very limited
Claude Sonnet	80–100	2.0–3.0s	Limited

Groq's custom LPU hardware serves GPT-OSS 120B roughly 5x faster than GPU-based APIs. For productivity tasks — summarizing, grammar, email drafting — output quality is equivalent to the best paid models, and the free tier needs no credit card.

04. The Prompt Engine

Every one of BEBO's 6 functions uses a structured 4-layer prompt instead of a naive instruction:

Temperature is tuned per task — deterministic where accuracy matters, creative where it helps:

Function	Temperature	Why
Fix Grammar	0.05	Fully deterministic — no creative drift
Summarize	0.20	Accurate, minimal variation
Draft Email	0.25	Structured, professional output
Simplify	0.30	Clear rewording with slight freedom
Ask AI	0.50	Balanced general answers
Humanize	0.65	Natural, creative variation

Every response also passes through a stripMarkdown() post-processor, so the output is always clean plain text — ready to paste anywhere without stray asterisks or formatting noise.

05. Global Shortcuts

Three system-wide hotkeys registered through Electron's globalShortcut API — they work inside any app: Word, Excel, your browser, anywhere on Windows.

Shortcut	Action
Ctrl + Shift + W	Wake up BEBO
Ctrl + Shift + B	Hide BEBO
Ctrl + Shift + P	Show / hide the assistant panel

New in BEBO 2.0: all three shortcuts can be remapped from the ⚙ settings inside the panel — pick any combo that suits your workflow. (Hide moved from Ctrl+Shift+H to Ctrl+Shift+B so Windows' Win + H voice typing stays untouched.)

06. Privacy by Architecture

BEBO is stateless by design. There is no account and no database. Your text travels directly from your machine to the Groq API over HTTPS and the result comes back — nothing is stored anywhere in between. The only things saved locally are your own API key and settings. The only thing BEBO ever reports is an optional anonymous daily ping — app version plus a random id, never your text — used purely to count active installs; switch it off anytime in the panel settings. The entire source code is public on GitHub, so none of this requires trust — you can read it.