Shadeform

We deploy fleets of AI Infrastructure Engineers to collaborate over week-long time horizons, run 100s of experiments in parallel, and learn from them to optimize your training/ inference infrastructure to your workloads. We've partnered with companies to speed up their GPU kernels to 6x SOTA, and improved the latency of voice models to twice the competition.

shared extracted concept terms: gpu, model; shared product terms: gpu, workload, deploy, model; shared product theme: developer infrastructure; same category: AI infrastructure; shared affiliation: Y Combinator

Overshoot makes it easy for developers to build and run real-time vision applications. AI can see and understand the physical world. This unlocks new applications in physical security, safety, gaming, robotics and general consumer products. Soon, video agents will watch your home and your pet when you’re away. However, existing platforms make it painful for developers to build these real-time applications. Slow inference. Limited model availability. Break at scale. Overshoot solves this. Today,

shared extracted concept terms: model, availability; shared product terms: easy, run, time, unlock; shared product theme: developer infrastructure; shared affiliation: Y Combinator

Expanse unlocks wasted GPU capacity. We recover idle compute through three capabilities: resource prediction (right-sizing job submissions before they reach the scheduler), optimisation suggestions (code and config changes researchers can apply themselves), and failure prediction (catching jobs that will fail before they consume hours of GPU time). We’re four engineers. We ran HPC and GPU training workloads at the largest quant funds and national supercomputing centres. We faced this problem fir

shared extracted concept terms: gpu; shared product terms: unlock, gpu, capacity, jobs; shared product theme: developer infrastructure; shared affiliation: Y Combinator

PoplarML lets you deploy any machine learning model to a fleet of GPUs as a ready-to-use and scalable API endpoint with one command.

shared extracted concept terms: model, api; shared product terms: deploy, any, model, gpus; shared product theme: developer infrastructure; same category: AI infrastructure; shared affiliation: Y Combinator

Cedana (YC S23) brings hyperscaler and frontier-lab orchestration capabilities for AI workflows. Our core capability is live migration for CPUs and GPUs workloads. This increases cost savings up to 80%, accelerates time to first token 2-10x, and enables stateful reliability of training jobs even through catastrophic GPU failures. We've integrated our solution into K8s, and support Kueue and Slurm for training distributed jobs, and Kserve for serving inference. OpenAI, Meta and Microsoft have fla

shared product terms: gpus, workload, cost, time; same category: AI infrastructure; shared affiliation: Y Combinator

Goodbye Slurm, Hello Konduktor. Trainy Konduktor is a software platform for AI teams to schedule workloads with priority, control resource allocation, and improve GPU reliability. With Konduktor, teams submit jobs to a healthy pool of GPUs, assign job priority with a simple user interface, and never worry about hardware faults again.

shared extracted concepts: gpus; shared product terms: gpu, workload, gpus, jobs; same category: AI infrastructure; shared affiliation: Y Combinator