Geometric
Geometric is an AI infrastructure company that automatically discovers and verifies CUDA kernel optimizations for machine learning workloads, delivering speedups as ready-to-merge GitHub pull requests. The platform analyzes model code and runtime behavior to identify bottlenecks, then generates and benchmarks optimized kernel implementations against agreed baselines. It also offers an inference API serving optimized open models like DeepSeek-V4.1-Flash, which Geometric's tools enabled to run on a single NVIDIA B300 at 260 output tokens per second.
- Developer Tools
- Artificial Intelligence
- Software Only
- Data & Analytics
Funding
Founders
Founder details are not available yet.
Product
Problem
Machine learning teams often underutilize their GPU infrastructure because identifying and implementing optimized CUDA kernels is a highly manual, time-intensive process that requires deep expertise. This results in significant wasted compute spend and slower model training and inference cycles, especially for teams operating at fleet scale.
Solution
Geometric provides an automated platform that discovers and verifies CUDA kernel optimizations for ML workloads. The system analyzes model code and runtime behavior to identify performance bottlenecks, then automatically generates optimized kernel implementations. These verified speedups are delivered directly as GitHub pull requests, allowing teams to integrate them into their codebase with minimal friction. By automating the entire optimization workflow, Geometric enables engineering teams to achieve substantial performance gains without dedicating specialized engineers to kernel development. The company also offers an inference API serving open models through its optimized deployments, and partnership engagements that span model quantization, kernel/runtime optimization, and deployment/serving tradeoffs.
Target Audience
Primary customers are machine learning engineering teams and platform teams at organizations deploying ML models at GPU-fleet scale, particularly those in AI infrastructure, large-scale model training, and high-throughput inference environments. Partnership engagements also target model teams, inference platforms, and application/device teams with specific performance constraints.
Features
- Automated CUDA kernel discovery that scans model codebases to identify suboptimal operations and replacement opportunities
- Verification pipeline that benchmarks proposed kernels against existing implementations to ensure correctness and measure actual speedup
- Direct GitHub pull request integration that packages verified optimizations as ready-to-merge code changes
- Fleet-scale analysis capabilities designed for organizations managing large GPU clusters and multiple ML models
- Performance monitoring that tracks speedup impact across training and inference workloads over time
- Inference API with OpenAI- and Anthropic-compatible endpoints, supporting models like deepseek/deepseek-v4.1-flash with a 1,048,576-token context length
- Partnership engagements covering model quantization, kernel/runtime optimization, and deployment/serving evaluation with agreed baselines and success criteria