About the Role
TLDR: Multimodal ML Engineer to train and ship vision, audio, video, and speech models for an AI safety platform that operates at 100M+ API calls/month.About usWhite Circle is an AI Safety company building the safety, reliability, and optimization layer for AI systems. At the core of our platform are policies – simple natural-language rules that define what an AI model should and shouldn’t do. We automatically test, enforce, and continuously improve these policies at scale.We’ve raised $11M from top funds, founders, and senior leaders at OpenAI, Anthropic, HuggingFace, Mistral, DeepMind, Datadog, Sentry, and othersWe process over 100M+ API calls every monthWe fine-tune and train our own LLMs so they run faster and cheaper than any open or proprietary modelWe’re a small, highly focused team. If you want to work deeply on hard problems, see your work ship to production quickly, and influence how AI safety is actually built – you’re the one we need.You willTrain and fine-tune large-scale multimodal models (vision-language, audio, speech) from scratch and from pretrained checkpointsExtend models across modalities: image understanding, video temporal modeling, long-context processing, and streaming audioDesign and run experiments: architecture changes, data mixes, training recipesBuild and maintain multimodal data pipelines — from raw images, video, and audio recordings to training-ready datasets, including synthetic data generationTrain and optimize MoE architectures for efficient multimodal inferenceBuild alignment pipelines: SFT, DPO, GRPO, reward modeling — across modalities, not just textOptimize models for production: quantization, distillation, batching, streaming and low-latency servingDeploy models end-to-end: from research checkpoint to production servingDefine evaluation metrics and benchmarks that actually matter for the product: visual QA, spatial reasoning, video comprehension, speech and audio understandingYou’ll fit right in if you3+ years training large-scale deep learning models in multimodal domains (vision-language, audio, speech, or acoustic)Strong PyTorch skills with hands-on distributed training experience (DeepSpeed, FSDP, or similar)Deep experience with multimodal architectures — you understand how vision/audio encoders, projectors, and LLMs fit together (LLaVA, Qwen-VL, InternVL, Audio Flamingo, Omni Qwen, Audio Qwen, Whisper, HuBERT, Conformer, or similar)Hands-on with RLHF/alignment for multimodal: GRPO, DPO, reward modeling — not just for textExperience with video and/or audio sequence modeling: temporal modeling, long-context processing, efficient attention, streaming inferenceTrack record of shipping models to production: you've hit latency targets and optimized inference, not just reported benchmark scoresComfortable with large-scale multimodal dataset curation: image-text pairs, video-instruction data, audio preprocessing, augmentation, synthetic data generationFamiliar with MoE architectures and their tradeoffs for multimodal workloadsStrong engineering fundamentals: clean code, version control, testing, documentationA big plus:Understanding of audio signal processing fundamentals (spectrograms, mel features, noise reduction)Why White CirclePaid time off in line with your local regulations, no matter where you work from