Multimodal ML Engineer

United States Digital Space LLC; Paris; Hybride; EUR 86 000 - 215 000; Plein temps

Avantages offerts par ce poste

Equity

Visa and relocation assistance

Résumé du poste

White Circle, an AI Safety company, is hiring a Multimodal ML Engineer to build policy enforcement and optimization layers for AI systems. You will train and fine-tune large-scale multimodal models across vision-language and audio modalities, and deploy production-ready pipelines.

Relocation to Paris or London (hybrid) is required. The role requires hands-on RLHF, MoE architectures, and experience with PyTorch, DeepSpeed, and large-scale data curation.

Qualifications

• 3+ years training large-scale multimodal models.
• Strong PyTorch and distributed training experience (DeepSpeed, FSDP).
• Deep familiarity with multimodal architectures – LLaVA, Qwen-VL, InternVL, Audio Flamingo, Whisper, HuBERT, Conformer or similar.
• Hands-on RLHF/alignment across modalities (GRPO, DPO, reward modeling).
• Both audio and video experience required – sequence modeling for each, plus large-scale dataset curation and production inference optimization.

Responsabilités

• Train and fine-tune large-scale multimodal models (vision-language, audio, speech, video) from scratch and from pretrained checkpoints.
• Design experiments, build multimodal data pipelines, and train MoE architectures.
• Build alignment pipelines (SFT, DPO, GRPO), optimize for production (quantization, distillation, streaming), and deploy end-to-end.
• Define evaluation metrics that actually matter for the product.

Connaissances

PyTorch

DeepSpeed

FSDP

RLHF

Multimodal models

MoE architectures

Data pipelines

Production inference

Outils

LLaVA

Qwen-VL

InternVL

Audio Flamingo

Whisper

HuBERT

Conformer

Description du poste

We're looking for a Multimodal ML Engineer to join White Circle, an AI Safety company building the policy enforcement and optimization layer for AI systems. Backed by $11M from senior leaders at OpenAI, Anthropic, HuggingFace, Mistral, and DeepMind, White Circle processes 100M+ API calls monthly and runs its own LLMs in production.

You will

• Train and fine-tune large-scale multimodal models (vision-language, audio, speech, video) from scratch and from pretrained checkpoints.

• Design experiments, build multimodal data pipelines, and train MoE architectures.

• Build alignment pipelines (SFT, DPO, GRPO), optimize for production (quantization, distillation, streaming), and deploy end-to-end.

• Define evaluation metrics that actually matter for the product.

Requirements

• 3+ years training large-scale multimodal models.

• Strong PyTorch and distributed training experience (DeepSpeed, FSDP).

• Deep familiarity with multimodal architectures – LLaVA, Qwen-VL, InternVL, Audio Flamingo, Whisper, HuBERT, Conformer or similar.

• Hands-on RLHF/alignment across modalities (GRPO, DPO, reward modeling).

• Both audio and video experience required – sequence modeling for each, plus large-scale dataset curation and production inference optimization.

• Relocation to Paris or London (hybrid) required.

Bonus

• Audio signal processing fundamentals – spectrograms, mel features, noise reduction.

• MoE architecture experience.

We offer

• $100k–$250k/year salary + equity; higher figures can be negotiated.

• Official employment, visa and relocation help.

Compensation: $100K – $250K Higher figures and equity are negotiable