Lei Zhou

Non-deterministic in, predictable out

Mixed criticality · Platform security · Physical AI

Lei Zhou

Principal Systems Architect & Technical Leader

Toronto area, Canada

Reliability, safety and security are engineered in, by design and by disciplined, traceable process: isolation, diverse redundancy and independence, demonstrated with measured evidence, never assumed.

Mission

Help teams run critical functions and AI on shared platforms, with behaviour they can predict, defend and evidence.

In the Physical AI era, cars, robots, factories and edge systems run real-time, safety- and security-critical workloads on the same silicon as non-deterministic noisy neighbours such as AI inference.

Safety and security both come down to three questions: Is the system intact? Is it behaving as intended? Can it recover to a safe, secure state when it isn’t?

My focus is designing the answers in, verifying them through system-level analysis, measurement and attestation, and building in resilience: fail-safe or fail-operational at runtime, secure recovery and incident response in the field.

Track record

What stands behind it

Predictable

Mixed-criticality and virtualisation
  • Conceived and led the real-time benchmark behind Linaro’s first Tier-1 automotive win
  • Optimised a virtio-gpu graphics pipeline to within the customer’s virtualisation-overhead target, at parity with the previous stack
  • Delivered industrial telemetry for national critical infrastructure, fail-operational over diverse redundant channels
  • Published MCIB, an open interference-benchmark specification; first measurement campaign planned

QNX · AGL · Android Automotive · Gunyah · Xen · KVM · timing analysis · interference attribution · GPU virtualisation

Secure

Full-stack platform security and confidential computing

Secure and measured boot · TrustZone · TEEs · attestation · device identity · key management · threat modelling · UNECE R155/R156

Accountable

Physical AI and trustworthy AI systems
  • Architected a sovereign LLM platform across three public clouds, with regional nodes delivered as proofs of concept
  • Built anchorstamp, a proof of concept that checks an AI agent’s claims against an independent observer

Robotics · sovereign AI infrastructure · AI agent accountability · independent evidence

About

How I got here

I’m a systems architect and player-coach technical leader for embedded, edge and cloud platforms where reliability, security and predictable performance are acceptance criteria, not aspirations.

My career spans industrial automation and telemetry at NARI; mobile-computing platforms at Qualcomm; platform and firmware security at AMD; OS security hardening and threat detection at BlackBerry; firmware security at NXM Labs; and platform-security delivery, SDV mixed-criticality virtualisation and sovereign AI infrastructure at Linaro. Today I continue with independent open research on Physical AI and agentic-AI accountability.

Mixed criticality in one picture: workloads that must never miss a deadline share a chip with workloads that only need to be fast. My work is keeping them apart, and measuring that they stay apart. Illustration, not measured data. For a working example on a PC, see vcockpit.
Selected work

What I’ve built and led

PredictableMixed-criticality and virtualisation

GPU virtualisation for a software-defined vehicle cockpit

Linaro · technical delivery lead and architect

A safety-relevant instrument cluster and an Android Automotive infotainment guest sharing one Arm SoC and one GPU under a type-1 hypervisor. Architected and led the guest graphics move from virgl to gfxstream (Vulkan, GLES via ANGLE, zero-copy rendering from guest memory over VirtIO-GPU), then the benchmark-led optimisation. Kept virtualisation loss within the customer’s target and reached parity with the previous stack.

Worst-case latency benchmark and outlier attribution for real-time Linux

Linaro · conceived and led

A benchmark framework for five RTOS primitives under PREEMPT_RT, using PMU and ftrace analysis to attribute latency outliers. It became Linaro’s first automotive benchmark capability and the technical foundation of its first Tier-1 automotive engagement.

SecureFull-stack platform security and confidential computing

Security and conformance assessment of an automotive SoC BSP

Linaro · assessment lead

Led the end-to-end assessment of an automotive SoC BSP and personally conducted the firmware vulnerability and security gap analysis: secure boot chain (integrity, authenticity, anti-rollback) and cryptographic architecture review, then a UNECE R155/R156 conformance gap analysis with a remediation roadmap.

TEE trusted services for an automotive telematics unit

Linaro · delivery lead

Hardened TLS with mutual authentication, secure storage and tamper-resistant logging, built as OP-TEE trusted services for a 5G telematics unit, toward UN R155 and ISO/SAE 21434 compliance. Delivered and accepted ahead of schedule.

Widevine L1 secure video playback (DRM) for in-flight entertainment systems

Linaro · delivery director and tech lead

Led delivery of NXP’s Yocto Linux Widevine L1 DRM solution on i.MX8MQ: an OP-TEE-based secure video path with secure boot. The product passed airworthiness qualification and Widevine L1 certification, the first commercial in-flight entertainment product on NXP’s Linux solution. That delivery led to further secure-boot and DRM projects, all of which I led.

IoT security platform, from requirements to certification

NXM Labs · Director of Firmware Security

Led a team of about six engineers and consultants from requirements to a commercial design win, owning the architecture and technical design reviews. Defined the security architecture, threat model and roadmap for attestation-gated device onboarding, with the security-sensitive control plane, audit logging and access control running inside a TEE secure enclave, and took the platform through PSA Certified Level 2 at an independent security laboratory.

AccountablePhysical AI and trustworthy AI systems

Sovereign AI infrastructure and AI-native engineering

Linaro · architect and delivery lead

Designed a provider-agnostic, three-tier sovereign LLM platform across three public clouds, owned its threat model and security controls, and personally delivered two of the regional nodes. Championed the company-wide move to AI-assisted engineering and ran a biweekly lunch-and-learn on hands-on practice.

Open research

Current projects

vcockpit: QNX and Linux cockpit displays on one PC

A shift-left development environment for vehicle cockpits: a QNX 8.0 instrument cluster and a Linux guest run in separate virtual machines and are composed into one display on a PC, with each frame handed over as a dma-buf file descriptor and never copied by the proxy, input forwarded to the guest under the pointer, and a one-command launcher. A separate monitor checks the telltales the driver would actually see. Entered in the QNX Global Hackathon 2026.

CQNX 8.0GPL-2.0-or-laterCC BY 4.0

MCIB: Mixed-Criticality Interference Benchmark

An open, platform-neutral specification for attributing interference between AI inference and real-time control on shared silicon: an interference taxonomy, an attribution protocol and a results schema with mandatory provenance. The harness is validated on a Raspberry Pi 4B under PREEMPT_RT; the first measurement campaign is planned on Renesas R-Car V4H.

Apache-2.0CC BY 4.0PREEMPT_RT

anchorstamp: attested execution evidence for AI agents

Checks what an AI agent claims it did against what an independent observer on the platform recorded, using signed, hash-chained records. The Linux kernel-audit proof of concept detects forged claims and unreported actions across 11 live scenarios.

RustMIT / Apache-2.0Spec part 1 published
Talks and writing

Speaking and publications

WCET outlier attribution for mixed-criticality automotive systems

Embedded World 2026, Nuremberg · AGL All Member Meeting 2026, Tokyo

Sovereign AI is a hardware problem

TAICO Toronto · July 2026

What deploying a sovereign LLM stack across three continents revealed about trust below the API layer, and why sovereignty needs hardware-rooted, attestable proof rather than policy alone.

Two talks at EdgeTech+ Japan 2025

EdgeTech+ Japan 2025

An RTOS determinism benchmark framework; and MRAM for compact Linux on resource-constrained safety systems.

Build Security & Privacy Compliant Embedded System

Linaro Connect London 2023 · co-presented with Loic Poulain

From regulation to silicon: requirements discovery from regulations and standards, threat modelling with STRIDE, the secure boot chain and its signing and provisioning infrastructure, TEE-backed key protection, security by isolation, protecting data at rest, in transit and in use, and keeping devices secure in the field with OTA updates and remote attestation.

Slides © Linaro, as presented; not covered by this site’s CC BY licence.
Career path

Where I’ve worked

  1. Independent researchPhysical AI and AI-agent accountability

    MCIB and anchorstamp, both open source.

  2. LinaroPrincipal Tech Lead

    Led a distributed team of senior engineers while serving as service team manager, pre-sales solution architect, delivery director and tech lead, for automotive, aviation and industrial customers.

  3. NXM LabsDirector of Firmware Security

    IoT security architecture, device identity and PSA Certified Level 2.

  4. BlackBerrySenior System Security Engineer

    TrustZone-based OS integrity monitoring; Android and Linux security hardening.

  5. AMDStaff Software Engineer

    Secure boot, root of trust and content protection for APU and GPU platforms.

  6. QualcommStaff Software Engineer

    Snapdragon display, multimedia and content-protection software, from kernel driver to SDK. Qualcomm SuperStar award.

  7. NARI GroupSystem Engineering Manager

    Early career in industrial automation: real-time telemetry and control for national critical infrastructure, where reliability was the acceptance criterion.

Master of Electronic & Electrical Engineering, Nanjing University of Science and Technology

Contact

Get in touch

I’m open to conversations about principal architect and technical leadership roles, and to consulting engagements in my focus areas. I’m based in the Toronto area and open to remote, hybrid or relocation for the right role. Email is the quickest way to reach me; my full resume is available on request.

How I can help as a consultant:

  • Predictable: mixed-criticality and virtualisation. Consolidation architecture reviews, interference and timing characterisation on real-time Linux and hypervisor platforms, and GPU virtualisation performance.
  • Secure: platform security. Secure boot, TEE and attestation architecture, BSP security assessments, and UNECE R155/R156 conformance gap analysis.
  • Accountable: Physical AI platform readiness. Characterising how AI inference interferes with real-time control on your silicon, and which countermeasures fit: tuning, partitioning or a hard ceiling.
About the mark

Non-deterministic in, predictable out

The Lei Zhou mark: orange dots converging through a sky-blue dot into a navy dot, inside a navy L-shaped frame

Safety engineering never promises zero failure. Security engineering never promises zero vulnerabilities. Now AI is acting in the physical world. So what can engineers actually promise?

The orange dots are the non-deterministic world: faults, adversaries, the unknown, and now AI inference. Some converge through the sky-blue point, where engineering happens: isolation, diverse redundancy, independence and measurement. They resolve into navy: predictable behaviour, within known limits. The navy L is the deterministic frame that holds it all, and the axes it is measured against.

Some dots never converge. That is residual risk, acknowledged and managed with evidence, not wished away. So we can’t promise that nothing will go wrong. We can promise that we know where it can go wrong, that we’ve contained it, and that our evidence doesn’t come from the system being judged.