Machine Learning Engineer

JOB TITLE

Senior Applied Machine Learning Engineer (audio / music generation)

We’re building an AI-powered music system focused on commercial-ready audio generation. Our initial priority is getting the music generation quality right – structure, musicality, consistency, and production readiness.

We are looking for a Senior Applied ML Engineer to own the end-to-end audio generation pipeline for our MVP. This role is hands‑on and pragmatic: you’ll fine‑tune open-source music models, integrate inference pipelines, and work closely with audio and backend engineers to deliver usable results quickly and efficiently. This role starts as a contract engagement (details below), with a path to full‑time position for the right fit.

• Terms:
  
  Fixed-term (5 months) | Potential full‑time conversion
• Compensation: $30,000 (Full 5 Month Term)
• Location:
  
  Hybrid/On‑site (Monrovia, CA)

• Fine‑tuning open‑source music generation models.
• Implement conditioning controls (beats per minute, key, mood, section, density).
• Training and deploying parameter‑efficient fine‑tunes (LoRA / adapters).
• Building reference‑conditioned generation.
• Support long‑form generation via chunking and continuation.
• Integrating with Backend inference pipelines and APIs.
• Collaborating with audio DSP engineers to ensure outputs are production ready.

• Strong experience with Python and PyTorch.
• Hands‑on experience with audio or speech generation models.
• Familiarity with diffusion or autoregressive generative models.
• Experience using or fine‑tuning open‑source ML models, familiar with HF Interfaces.
• Understanding of audio representations.
• Experience deploying ML models to production or API environments.

• Familiarity with CLAP / audio embeddings or retrieval‑assisted generation.
• Experience working with LoRA / PEFT methods.
• Basic understanding of audio production workflows (tempo, key, stems, loudness).
• Experience Optimizing inference cost and latency.

• Reliably generate musically coherent, commercial‑friendly cues (30 ~ 120 seconds)
• The model responds correctly to conditioning inputs like tempo, key and mood
• Outputs are stable, repeatable and usable downstream by post‑production tools
• The system is modular and ready to be integrated with downstream models.