Liquid AI builds optimized AI systems using low-level GPU infrastructure (CUDA, custom kernels, TensorRT) paired with production training frameworks (PyTorch, DeepSpeed, Megatron-LM). The stack reveals an engineering-first org solving performance constraints: projects span custom CUDA kernel development, data-loading bottleneck elimination, and on-device VLM inference under latency constraints. Adopting llama.cpp and ExecuTorch signals focus on efficient edge deployment alongside enterprise-scale training.
Tech Stack 44 technologies
Core StackPyTorch Python TensorFlow C++ Rust Hugging Face Slack CUDA Nsight Systems Nsight Compute cuDNN cuBLAS C/C++ vLLM DeepSpeed Megatron-LM NCCL JAX GPU SGLang TensorRT Jupyter Discord iOS Android Jetson llama.cpp ONNX executorch Qwen+11 more
Adoptingllama.cpp ExecuTorch
What Liquid AI Is Building
◆Challenges
- Low latency
- Adapting vlms for enterprise customers
- Scaling rapidly
- Generic frameworks insufficient for performance
- Catching and fixing edge cases where llm may fail
- Ensuring datasets meet enterprise-grade quality
- Improving training throughput
- Reducing training cost
- Eliminating data loading bottlenecks
- Running vlms on-device under real-time constraints
▲Active Projects
- Build and scale data pipelines for audio model training
- Reusable applied workflows and tooling
- Custom cuda kernel development for novel model architectures
- Demo and interactive experience development
- Benchmark harness development for performance regression
- Japanese dataset curation and augmentation
- Japanese llm fine-tuning for enterprise use cases
- Evaluation framework implementation for japanese datasets
- Distributed training infrastructure for gpu clusters
- Data loading systems for multimodal datasets
Hiring Activity
20 roles · 10 in 30d
Department
Seniority
Company intelligence
