Natalie Serrino is a co-founder of Gimlet Labs and an engineering leader behind KForge, a system that uses AI to generate and optimize the specialized kernels powering machine-learning workloads across different processors. Her work tackles a central obstacle to efficient AI infrastructure: applications increasingly span multiple models, tools, and computational stages, yet adapting their software to unfamiliar accelerators requires scarce systems expertise.
Serrino studied computer engineering at Brown University, where she helped establish Spira, an engineering program for high-school girls supported by the National Science Foundation. She subsequently became a founding engineer at Observe and Pixie Labs. Following Pixie’s acquisition by New Relic in 2020, she worked there as a principal engineer, developing infrastructure observability techniques including Kubernetes autoscaling based on application-specific HTTP traffic.
At Gimlet, which launched publicly in October 2025, Serrino applies that operational background to heterogeneous inference orchestration: assigning individual stages of agentic applications to processors suited to their distinct computational and memory requirements.
- KForge and cross-platform kernel generation: Serrino co-authored the research introducing KForge, which combines code generation, compilation feedback, correctness checks, and profiling to optimize workloads for NVIDIA CUDA and Apple Metal. A subsequent study extended the approach to NVIDIA B200 and Intel Arc B580, reporting a 2.12% end-to-end throughput improvement on the B200 and a 5.13-fold geometric-mean speedup across a defined Intel benchmark subset.
- Optimization above and below the kernel: Serrino’s technical demonstrations show agents fusing multiple operations, rewriting one-dimensional average pooling as convolution to exploit stronger Metal implementations, and simplifying computation directly in PyTorch. She also identifies hard limits: agent-generated matrix multiplication can lose badly to carefully optimized existing libraries, while dramatic benchmark gains may reflect exploitable test inputs instead of genuine improvements.
- Hardware-in-the-loop verification: Her optimization workflow evaluates generated code on actual processors using representative inputs, floating-point tolerances, cache controls, reliable execution timing, and human supervision. She also co-authored work on formally verifying generated GPU kernels that translates implementations into a shared representation and uses the Z3 solver to establish equivalence or produce counterexamples.
Serrino treats generated kernels as a force multiplier for systems specialists: useful for exploring optimizations and porting workloads across hardware, provided every claimed improvement withstands correctness checks and real-world measurement.
