Own our AI models end-to-end: dataset strategy, training/fine-tuning, evaluation, optimization, and deployment to real hardware in the field. You’ll turn messy real-world streams into robust, low-latency inference pipelines.

What you’ll do

• Design, train, and fine-tune text/audio/vision models (e.g., DistilRoBERTa, wav2vec2, YOLO) for threat and aggression detection.
• Build reproducible training pipelines (HF/PyTorch/TF/SpeechBrain), incl. PEFT/LoRA, adapters, and transfer learning.
• Optimize for real-time: quantization, pruning, ONNX/TensorRT, mixed precision, batching, caching.
• Ship models to edge & cloud with CI/CD, versioning, and rollback; instrument latency and accuracy SLAs.
• Create data pipelines: collection, labeling, augmentation/synthesis, dataset versioning (DVC).
• Collaborate with backend/infra on streaming, Pub/Sub, autoscaling, and observability.

Must haves

• Python wizardry and strong PyTorch.
• Hands-on experience with Audio/Video AI - HuggingFace Transformers, SpeechBrain or torchaudio, one detection/pose stack (YOLO, MediaPipe).
• Production MLOps: experiment tracking, model registries, CI/CD, model monitoring.
• Comfortable with GCP, containers, and GPUs.
• Located in London or ready to relocate.

Nice to haves

• Real-time inference experience
• Streaming systems, low-latency audio (VAD/Whisper/wav2vec2), and CV pipelines.