L
Liquid AI is a GPU-focused AI infrastructure company built on CUDA, PyTorch, and low-level kernel optimization (custom CUDA kernels, TensorRT, vLLM). The tech stack and project list reveal a shift away from generic frameworks toward hand-tuned performance: they're developing custom inference kernels for edge hardware, optimizing RL training runs, and integrating novel optimization techniques—all pain points rooted in generic frameworks being insufficient for production scale. The engineering-heavy org (14 senior engineers, 6 hires in the last month) and active Japanese language model work suggest both rapid infrastructure development and vertical market expansion.
Tech Stack 34 technologies
Core StackPyTorch Python Hugging Face React Next.js TypeScript Tailwind CSS Supabase TensorFlow FastAPI C++ CUDA Nsight Systems Nsight Compute cuDNN cuBLAS C/C++ DeepSpeed Megatron-LM vLLM NCCL shadcn/ui Neon Apollo GPU JAX SGLang llama.cpp TensorRT Google Workspace+1 more
What Liquid AI Is Building
◆Challenges
- Scaling rapidly
- Generic frameworks insufficient for performance
- Eliminating manual invoicing workflows
- Dataset scarcity
- Adapting foundation models to new languages
- Catching and fixing edge cases where llm may fail
- Ensuring datasets meet enterprise-grade quality
- Improving training throughput
- Scaling data pipelines
- Eliminating data loading bottlenecks
▲Active Projects
- Rl training run optimization
- Pretraining data pipeline
- Synthetic data generation pipeline
- Optimization technique integration
- Custom cuda kernel development for novel model architectures
- Implement inference kernels for edge hardware
- Web crawler for dataset acquisition
- Japanese dataset curation and augmentation
- Japanese llm fine-tuning for enterprise use cases
- Evaluation framework implementation for japanese datasets
Hiring Activity
Steady20 roles · 6 in 30d
Department
Seniority
Company intelligence
