Build what’s next.
At Ample AI.

An AI model that optimizes inference, not an army of engineers.

Today's inference providers grew by putting forward deployed engineers on every customer. RSI² observes each workload, optimizes every layer from the agent layer to the hardware kernel, and validates each change through simulation, replay, and canary before it ships.

It learns from expert feedback, verifiable rewards, and AI self-play, so every new problem makes the platform better. It runs across NVIDIA, AMD, Google TPU, and Intel, and on frontier APIs as well as open models.

Inference already outspends databases.

Annual inference spend has passed $200B in three years and is tracking toward $1T. Agentic and reasoning workloads need 10-100x more compute per query, making inference a CFO-visible line item. 5,000+ companies now run their own inference, and almost none can build an optimized full-stack setup in house.

Most of that spend goes to OpenAI, Anthropic, and about six inference providers. Chips and models keep fragmenting, and that complexity is exactly what Ample automates.

One platform, three ways to buy.

Autonomous Optimization is a managed service that sits on top of any existing stack and shows value in days. Workload Optimized Inference deploys on neocloud, hyperscaler, or customer-owned servers. Integrated Inference Systems pair co-designed hardware and software for the most tokens per dollar and per watt.

Sellers land customers directly and co-sell with chip, server, and cloud partners including AMD, Google TPU, Dell, and HPE.

A full-stack, multi-cloud vision.

Ample’s unified platform spans clouds and hardware topologies. Its recursive self-improvement engine is designed to replace manual tuning with a system that learns, optimizes, and improves with each execution run.

Capacity and commercial expansion.

Ample expects about $100M of inference capacity to come online around March 1. The commercial plan is to land 10–15 customers per segment with the founding team, then scale the sales organization around that playbook.

Mountain View, California. Hybrid.

The three Scalrr searches are based in Mountain View on a hybrid basis. Ample is growing from a small founding team toward roughly 25–30 employees over the next year, giving early hires substantial ownership and influence over the technology and company.

Ursheet Parikh

Co-Founder & CEO

• Co-founded StorSimple, acquired by Microsoft in 2012 and scaled into a meaningful part of Azure
• 10 years as a partner at Mayfield investing in AI and cloud infrastructure, including Rancher Labs and CloudGenix
• Built a new business from zero to scale inside Cisco

Rajat Monga

Co-Founder & CTO

• Co-created TensorFlow and led it through 2019 as a founding member of Google Brain
• Built eBay's search engine and scaled ad serving at Google
• Founded Inference.io, then led AI Frameworks at Microsoft

Priya Gupta

Founding Engineer

Amit Aggarwal

VP Engineering

Estimated base salary

$225K-$350K, plus significant early-stage equity. Final compensation is discussed with the team.

Early-stage equity

Join a team of roughly a dozen people, early enough for your ownership and technical decisions to matter.

Work with the founding team

Build alongside TensorFlow co-founder Rajat Monga and collaborate across research, inference infrastructure, and systems engineering.

Hard problems, real ownership

Help define core architecture across model optimization, distributed inference, compilers, and accelerator software.

What you’ll do

• Design, build, and automate the large-scale distributed systems that serve production models globally.
• Develop intelligent request routing, load balancing, and traffic management systems across thousands of accelerators.
• Implement LLM inference optimizations, including advanced batching, caching, disaggregation and fleet-wide request management strategies.
• Maximize compute efficiency across the fleet by autoscaling and orchestrating production, research, and experimental workloads.
• Build and automate production-grade deployment pipelines using Kubernetes and cloud infrastructure (AWS, GCP, Azure).
• Provide high-performance inference infrastructure that enables researchers to rapidly prototype and validate new model architectures.
• Integrate new AI accelerator platforms and adapt systems in real time to live traffic patterns and system failures.

Who you are

• Significant software engineering experience with high-performance, large-scale distributed systems.
• Direct experience implementing and deploying machine learning systems at scale.
• Hands-on experience with load balancing, request routing, or traffic management systems.
• Familiarity with LLM inference optimization techniques, batching algorithms, and caching strategies.
• Production experience with Kubernetes and major cloud infrastructure (AWS, GCP, Azure).
• Proficiency in Python, Rust, or C++.
• Results-oriented mindset with a bias toward flexibility, high impact, and strong technical ownership.

What you’ll do

• Develop novel post-training and alignment algorithms for frontier language models
• Design and improve training pipelines using techniques including SFT, DPO, RLHF, RLVR, GRPO and preference optimization
• Research model compression, speculative decoding, draft models and efficient inference techniques
• Collaborate closely with inference infrastructure engineers to translate research into production systems
• Improve model quality while reducing latency, compute cost and power consumption
• Publish novel ideas internally and rapidly iterate from research into deployed systems

Who you are

• PhD or equivalent industry experience in Machine Learning, AI or a related discipline
• Strong publication record or demonstrated research excellence
• Experience training or fine-tuning large language models
• Deep understanding of transformer architectures and optimization
• Hands-on experience with one or more: / RLHF / DPO / GRPO / RLVR / Reward Modeling / SFT
• Strong Python and PyTorch experience
• Comfortable working across research and production engineering

What you’ll do

• Design and optimize high-performance CUDA, HIP, PTX, and Triton kernels for large-scale LLM inference.
• Build fused operators, attention kernels, GEMM optimizations, and memory-efficient execution paths for modern transformer architectures.
• Improve performance through advanced GPU optimization techniques including shared memory utilization, Tensor Core acceleration, memory coalescing, vectorization, communication/computation overlap, and asynchronous execution.
• Collaborate across compiler, runtime, and hardware layers including MLIR, LLVM, XLA, Triton, CUDA, and custom accelerator software stacks.
• Profile, benchmark, and eliminate bottlenecks across compute, memory bandwidth, communication, and kernel execution.
• Partner closely with inference and distributed systems engineers to optimize continuous batching, KV-cache management, quantization, and heterogeneous hardware scheduling.
• Help shape the software architecture supporting next-generation AI accelerators beyond NVIDIA GPUs.

Who you are

• Expert-level C++ development with CUDA, Triton, HIP, PTX, or other GPU programming environments.
• Experience building or optimizing GPU kernels, AI accelerator software, compiler backends, or high-performance numerical libraries.
• Deep understanding of GPU architecture, memory hierarchy, SIMD/SIMT programming, Tensor Cores, cache behavior, and parallel programming.
• Experience with one or more compiler or runtime technologies such as MLIR, LLVM, XLA, Triton, CUTLASS/CuTe, TVM, or TensorRT-LLM.
• Familiarity with modern LLM inference systems including FlashAttention, FlashInfer, vLLM, SGLang, PagedAttention, TensorRT-LLM, or similar technologies.
• Strong Linux systems programming and performance profiling experience.
• Experience developing production systems software for AI infrastructure, HPC, distributed computing, or accelerator platforms.

Scaling TensorFlow, Navigating Startup Pivots, ML Edge Infrastructure and AI Inference Strategy

The Engineering Leadership Podcast · Apr 2026

Rajat on scaling TensorFlow and where AI inference is headed.

A Conversation with Rajat Monga: From TensorFlow to Next-Gen AI

The Effortless Podcast

TensorFlow, the rise of PyTorch, and AI inferencing.

Ursheet Parikh

Mayfield

Ursheet's operating and investing background, from StorSimple to Mayfield.

Ursheet Parikh's Cradle-to-Exit Role at Rancher Labs

Mayfield · Mar 2021

How Ursheet partnered with Rancher Labs from its early days to its sale to SUSE.