Private AI Assistant

Boumb'App features a scale-constrained melodic AI generator powered by a dedicated self-hosted small language model (SLM). It transforms natural language prompt instructions into ready-to-play musical sequences.

1. Privacy & Infrastructure Philosophy

Unlike solutions relying on commercial cloud APIs (OpenAI, Google Cloud, AWS):

• Inferences run on our private, dedicated Linux server located in France.
• Zero personal data retention: Prompts and musical notes are processed strictly anonymously without IP association or user account coupling.
• Zero data mining: Your musical ideas are never harvested to train third-party commercial models.

2. Inferences & Processing Pipeline

Melodic generation flows through several synchronized verification layers:

User prompt ("energetic trap bass in D minor")
      ↓
Style Classifier / Intent Detector
      ↓
Composer Model (Custom Qwen 1.5B quantized GGUF)
      ↓
Deterministic Harmonic Guard
      ↓
Direct Note Injection into Piano Roll

Deterministic Harmonic Guard

Generative models can occasionally produce out-of-scale anomalies. To maintain strict musical usability, our deterministic Harmonic Guard inspects every generated pitch against your project's active musical scale and automatically corrects dissonant notes without modifying rhythmic cadence.

3. Effective Prompting Guidelines

The assistant parses natural language in English and French, recognizing standard beatmaker and audio production terminology:

Recommended Examples

• punchy electro bassline in D minor
• warm neo-soul chord progression in A minor
• fast energetic synthwave arpeggio with high density
• melancholic lo-fi keys in F# dorian
• dark rolling techno bassline 16 steps

Recognized Prompt Parameters

• Genre & Feel: Trap, Lo-fi, Funk, Synthwave, Techno, Ambient, Jazz, etc.
• Instrument Role: bass, lead, chords, arpeggio.
• Musical Key: Specifying a key in the text (e.g. in C# minor) overrides the current global TopBar key for that specific generation.

4. Concurrency Limits & Queue Management

To ensure equitable computing distribution without server degradation:

• Processing capacity is capped at 3 concurrent generation requests.
• If capacity is temporarily reached, a queue mechanism holds requests for up to 15 seconds.
• In rare peak traffic scenarios where the queue expires, the interface reports AI service temporarily unavailable, while the rest of the studio remains 100% operational offline.