Chenyuan Qu

Enterprise AI · Computer vision research

Chenyuan Qu

I build AI tools that help people get real work done, and I research new ways for computers to understand and create visual content.

At Allsee, I work with customers and teams to turn business problems into software used across the company and by customers around the world. I still spend much of my time building—from Python services and device software to AI tools for everyday operations. Alongside this, I am a part-time PhD researcher at the University of Birmingham.

Google ScholarGitHubLinkedInORCIDEmail

University of Birmingham · Allsee · Vieunite

Portrait of Chenyuan Qu

Enterprise AI

See industry work
200,000+managed devices
Serving 1,000+ organisations and 50,000+ users around the world.
≈80%less time per task
Less time needed to complete the ERP tasks that were tested.
≈99%fewer human errors
Fewer human mistakes in the ERP tasks that were tested.

Research

See research
BMVC 2025first author
VisualSplit explores image representations people can understand and edit.
CVPR 2024oral paper
Co-author of 360+x, selected for an oral presentation.
4 paperspeer reviewed
Published at BMVC, CVPR, ICASSP, and ICCV.

Enterprise AI

Turning real business problems into useful AI.

My work usually starts by speaking with customers or colleagues, understanding where time is being lost, and then building and improving a tool until it works well in day-to-day use.

Nexus MySignagePortal

Allsee's previous platform had grown into around ten different versions. Nexus replaced them with one shared platform that is easier for customers to use and for the team to improve.

What I did

I began by visiting customers and watching how they worked. I then led the redesign, wrote most of the Python/FastAPI backend, contributed to the Java software on the devices, and guided a five-person team.

What changed

Nexus now manages more than 200,000 devices for 1,000+ organisations and 50,000+ users worldwide. Larger features that used to take about a month can often be delivered in a few days, monthly reported bugs fell from dozens to low single digits, and the platform helped win at least £50,000 in directly attributable sales.

How it works

Tools and skills

  • Customer research
  • Python / FastAPI
  • Java
  • AWS
  • Device software
  • Team leadership

An AI assistant for everyday ERP work

Employees can describe an order, warehouse, or repair task in everyday language and let the system complete the approved steps inside the company's ERP.

What I did

I built the AI layer. It checks what each person is allowed to do, asks for confirmation before sensitive changes, and records every action. Behind the scenes, reusable MCP tools connect the model to the ERP without allowing it to bypass normal business rules.

What changed

In the day-to-day tasks we tested, employees completed the work around 80% faster and made around 99% fewer manual mistakes.

How it works

Tools and skills

  • AI agents
  • MCP
  • Permissions
  • Safeguards
  • Human approval

COMPaD

Most image generators leave a finished picture that is hard to change. COMPaD aims to create a real design file, so text, product images, logos, shapes, and QR codes can still be edited.

What I did

I started and lead this joint project between Allsee and the University of Birmingham. I designed and trained a model that works with both text and images, built on open-source Qwen2.5. I also built the agents and online services that break a brief into smaller tasks and keep the design information organised as the system works.

Where it is now

The system is in first-stage alpha testing. We check whether the poster looks right, whether each element is placed correctly, and whether the final result is genuinely editable.

How it works

Tools and skills

  • Model training
  • Qwen2.5
  • AI agents
  • Editable output
  • Testing
  • Deployment

Research questions

The ideas that connect my current work.

01

Making image features easier to understand

I separate shape, colour, and brightness so we can see—and edit—what an AI model is using.

Read VisualSplit02

Understanding a scene from more than one viewpoint

I combine panoramic and first-person views with sound, text, and location to give AI a fuller picture of a place.

Read 360+x03

Giving people more control over generated images

I explore how generative models can rebuild, restore, and edit images without hiding every decision from the user.

Read VisualSplit

Research

Selected research

I want AI systems to understand images without making every decision invisible to people. These two projects show how I approach that question from image representation and from richer, multi-view data.

BMVC 2025 · Image representation

VisualSplit

VisualSplit asks whether an AI model can understand and rebuild an image using three simple parts: edges, colour regions, and overall brightness.

BMVC · First-author research2025

Main idea

When shape, colour, and brightness stay separate, it becomes easier to understand what the model sees and to change one part without changing everything else.

Question
Modern vision models often mix shape, colour, and lighting into features that people cannot easily inspect or control.
My role
As first author, I developed and evaluated the approach with my co-authors. We tested how the three inputs could rebuild images and support editing, restoration, and image generation.
Where published
The work was published at BMVC 2025. The paper, talk, code, model weights, and working examples are all public.
ProjectPaperCodeModels

CVPR 2024 · Oral paper

360+x

360+x lets researchers study the same scene through several camera views, together with sound, location, and text.

CVPR Oral · Collaborative research2024

Main idea

A scene makes more sense when an AI system can connect what is visible from different viewpoints with what can be heard and where it happened.

Question
Most datasets show a place through one camera or one type of data. That makes it difficult to study how different views and signals relate to one another.
My role
I was one of six authors. We built the dataset, research benchmarks, and publication together as a team.
Where published
The paper was selected for an oral presentation at CVPR 2024. The paper, code, dataset, poster, and project video are public.
ProjectPaperCodeDataset

Papers & datasets

Publications

My peer-reviewed papers, with the code, datasets, figures, talks, and citation details available for anyone who wants to go deeper.

2025
2024
2023

Datasets

2025

BinEgo-360

A research dataset and challenge combining stereo first-person video, 360° views, spatial sound, text, and location information.

ChallengeRepository

2026

text-to-art-database

A privacy-safe text-to-image dataset on Hugging Face, organised so researchers can load and work with the images more easily.

Dataset

Experience

Experience & education

The work and study that shaped how I build, lead, and research.

Research appointments

University of Birmingham · MI X Group

Feb 2023 — Present
PhD Researcher (part-time)
Sep 2023 — Expected 2028

I study how AI can understand and create visual content in ways that are easier to interpret, combine, and control.

Research Assistant (part-time)
Dec 2023 — Present

I research how foundation models represent and combine visual ideas. This work includes COMPaD, the joint Allsee–University of Birmingham project for editable poster generation.

Research Assistant (part-time)
Feb 2023 — Dec 2023

I worked with domain researchers to apply machine learning to hydrology while keeping the results understandable to scientific users.

Education

Master's in Artificial Intelligence and Machine Learning

University of Birmingham

2021 — 2022

Graduated with Distinction, then continued at Birmingham for research work and a part-time PhD.

2018 — 2021

Physics gave me a strong base in mathematics, modelling, and experimental thinking that I still use in engineering and research.

Industry experience

Allsee · Vieunite

Sep 2022 — Present
Head of Technologies
Dec 2024 — Present

I lead the company's technology work while continuing to build software myself. My focus includes customer platforms, device software, the ERP, and practical AI tools for everyday work.

More about this role
  • Led Nexus from customer visits and early design through launch, while writing most of its Python/FastAPI backend and contributing to the Java device software.
  • Led the replacement of the company's ERP and built an AI layer that lets staff complete approved tasks in everyday language.
  • Started and lead COMPaD, a joint industry–university project for creating editable commercial posters with AI.
  • Work directly with customers, sales, operations, and engineers to decide what should be built and to explain the trade-offs clearly.
Full-stack Engineer
Dec 2023 — Dec 2024

I worked across the customer platform, ERP, cloud services, and device software, helping replace a collection of hardware-specific systems with one shared product.

More about this role
  • Helped turn around ten separate platform versions into one codebase that could still work with the existing device range.
  • Built core parts of the replacement ERP and connected it with the CMS, Salesforce, and QuickBooks.
  • Worked across the user interface, online services, software releases, and device constraints instead of treating them as separate problems.
Algorithm Engineer
Sep 2022 — Dec 2023

I built backend services, recommendation features, and generative-AI tools for Allsee and Vieunite, including features used in the Vieutopia AI art product.

More about this role
  • Connected AI features to the online services and day-to-day processes needed by the product.
  • Built recommendation features for content and product discovery.
  • Developed image-generation features that people could use through Vieutopia.

AsiaInfo Software Co. Ltd

Jul 2020 — Sep 2020
Algorithm Engineer Intern
Jul 2020 — Sep 2020

I built and improved machine-learning features for a virtual customer-service presenter designed to run on mobile devices.

News

Recent updates

Recent milestones from my work, research, and professional development.

COMPaD entered first-stage alpha testing

The current tests look at three things: whether the generated poster works visually, whether its layout is understood correctly, and whether the final file remains editable.

Began Help to Grow: Management at BCU

I joined the 12-week course at Birmingham City University Business School to strengthen my practical knowledge of strategy, digital change, marketing, operations, finance, and growth.

Read more

VisualSplit accepted to BMVC 2025

I released the paper, talk, code, model weights, and examples so the work can be read, tested, and reused.

Read more
Earlier updates

DIFF presented at ICASSP 2025

Our team showed how features from image-generation models can help a segmentation system work across different visual domains.

Read more

360+x selected for a CVPR 2024 oral presentation

The paper, code, research benchmarks, and dataset are available through the public project page.

Read more

Started my PhD in the MI X group

I began my part-time PhD in the University of Birmingham's MI X Group, studying how AI can understand images, sound, text, and different viewpoints together.

Read more

MeD published at ICCV 2023

This collaborative project learns to remove image noise by comparing several noisy views, without relying on clean target images.

Read more

Contact

If you would like to talk about applied AI, computer vision, or a potential role, email is the easiest way to reach me.

Chenyuan.Qu@outlook.com

Other addresses

Allsee
henry.qu@allsee-tech.com
Vieunite
henry.qu@vieunite.com
University of Birmingham
cxq134@student.bham.ac.uk