Skip to content
AGIACC.

AGIACC · Company brief

Trustworthiness-native AI infrastructure.

From chips and new devices to system software and applications. Security, execution safety and efficiency, designed together.

AGIACC is a startup founded by Dr. Xiaoyang Sun to develop secure, efficient computing infrastructure for AI agents and enterprise AI. Our ambition is to make access limits, memory protection and failure handling part of the computing stack, with evidence of how those protections affect useful work and resource cost.

Why this opportunity matters now
#

AI agents can execute code, use credentials and interact with enterprise services. Their ability to act creates infrastructure requirements: authority must be bounded, faults contained and resource use controlled. Training and inference add demanding memory, scheduling and compute requirements. Protection and efficiency must work together for deployment to be practical.

Industry developments support this direction. NVIDIA OpenShell brings explicit runtime policies to agent execution. Apple’s Memory Integrity Enforcement combines hardware and operating-system protections for memory safety. We see an opportunity to connect protection across layers and make its behavior and costs testable.

Founder experience that fits the opportunity
#

Xiaoyang earned his PhD at the University of Leeds and has conducted research at Leeds and the University of Glasgow. His industry experience includes R&D internships at Alibaba; he also consults on AI infrastructure at Kuaishou Technology.

Industry collaboration

Experience with operational AI systems

Xiaoyang co-authored KAIR, a distributed-training diagnosis study evaluated in Kuaishou production clusters.

One architectural thesis across the stack

An application’s permissions, a runtime’s isolation boundary and a device’s memory protections govern different resources. Our thesis is that designing their interfaces together can make execution limits more meaningful and reduce the cost of integrating them. The long-term scope spans three connected layers:

Long-term scope

  1. Hardware and new devices

    Chip and device co-design for memory protection and controlled resource access, with performance and energy considered from the outset.

  2. System software

    Isolation, permission enforcement, fault handling and resource management across agent execution, AI training and inference.

  3. Application frameworks and services

    Interfaces and platforms that make underlying protections usable, with inspectable behavior and measurable operating costs.

Security and execution safety · Resource efficiency · Reproducible evaluation

The intended advantage is the combination: prevent specified invalid operations, contain faults or compromise, and preserve practical performance.

Who we aim to serve
#

Our intended users are platform, security and AI infrastructure teams deploying AI with sensitive data, executable tools and enterprise services. The goal is to help them establish what an AI component may access, how failures are handled and what protection costs in latency, memory and throughput. Enterprise agents provide a starting customer focus; industrial operations, robotics and embedded AI extend the opportunity to devices, timing and recovery.

Commercial direction and seed purpose
#

Our current foundation is research and technical artifacts. Seed investment would advance integrated engineering, reproducible protection and performance evidence, and customer evaluations.

The initial commercial hypothesis is paid technical evaluation and integration, followed by supported software, platforms and services as demand becomes clearer. Deeper chip and device co-design would follow demonstrated requirements and suitable partners. The roadmap details the delivery stages and evidence required to advance.

We welcome seed investors, prospective design partners and researchers who share the ambition to make trustworthiness a foundation for capable AI.

Start a conversation

contact@agiacc.com

AGIACC · agiacc.com · Updated