PyTorch & JAX
Primary research frameworks — PyTorch for production flexibility, JAX for high-performance custom gradient computation.
Hugging Face
Transformers ecosystem for NLP foundation models — with our custom tokenisers, heads, and precision fine-tuning toolchain.
TensorRT & ONNX
Production inference optimisation — hardware-specific kernel fusion, layer optimisation, and mixed-precision compilation.
Weights & Biases
Experiment tracking, hyperparameter sweeps, model versioning, and real-time training monitoring for every project.
Triton & CUDA
Custom GPU kernels for attention variants, sparse operations, and hardware-specific memory-bandwidth-optimised inference.
Ray & DeepSpeed
Distributed training infrastructure — ZeRO optimisation, tensor/pipeline parallelism, and gradient checkpointing for large models.
TensorFlow Lite & CoreML
Edge deployment compilation targeting mobile, embedded, and hardware neural engine targets with latency benchmarking.
Foolbox & ART
Adversarial robustness evaluation and certified defence implementation — systematic red-teaming of every production model.