Skip to content

Vladislav Li

Email: li.vlad.viktor@gmail.com | Location: London, UK | GitHub: couper64 | Portfolio: vladislav.li

Profile

ML Infrastructure Engineer and PhD graduate in Artificial Intelligence with five years of hands-on experience building, deploying, and operating ML systems in production. Specialises in GPU server infrastructure, containerised ML service design, model serving and inference optimisation, and the full deployment stack from bare-metal Linux setup through to cloud and Kubernetes-based orchestration. Sits at the intersection of ML engineering and infrastructure: comfortable configuring hardware, deploying distributed backend services, and keeping production ML systems reliable, observable, and efficient. Published across IEEE, Image and Vision Computing, and Journal of Systems Architecture with 75+ citations.

Skills

ML Infrastructure & MLOps

  • Docker and Docker Compose — multi-stage builds, containerised ML service design, multi-service orchestration
  • Kubernetes — kubeadm cluster setup; pod, deployment, service, and ingress management (actively deepening via pet project)
  • CI/CD — GitHub Actions: automated test, build, and deployment pipelines; multi-repo synchronisation
  • ML model serving — inference pipeline design, GPU workload management, throughput and latency optimisation
  • ONNX Runtime — production real-time inference (50 Hz locomotion policy execution)
  • GPU server infrastructure — bare-metal setup, driver configuration, workload management, hardware monitoring
  • Experiment tracking — Weights & Biases, MLflow
  • Grafana — infrastructure and ML service monitoring dashboards
  • OpenResty — Lua-scripted reverse proxy for production traffic routing

ML & Computer Vision

  • PyTorch, TensorFlow/Keras — model training, fine-tuning, and evaluation
  • Object detection — YOLOv3, YOLOv5, YOLOv9, YOLOX, RT-DETR, Faster R-CNN, CenterNet, RetinaNet; few-shot detection and data augmentation strategies
  • Multi-object tracking — ByteTrack; real-time person tracking with stable ID assignment
  • Generative models — StyleGAN2, Stable Diffusion 1.5 with LoRA fine-tuning, TimeGAN (time-series synthesis), ESRGAN (super resolution)
  • Model distillation — PatchTST, MobileNetV3→MobileNetV2; wrapping partner pipelines in production stacks
  • Continuous / lifelong learning — image and time-series domains
  • OpenCV — image processing, augmentation, and evaluation pipelines
  • Few-shot learning — survey-level depth; energy-efficient inference for industrial edge settings

Robotics & Simulation

  • MuJoCo — physics simulation; sim-to-real pipeline with HAL abstraction
  • Behavior Trees — py-trees; reactive robot coordination, task sequencing, failure recovery
  • CycloneDDS — real-time publish/subscribe middleware for hardware robot communication
  • MCAP + Rerun — session recording, replay, and multi-stream synchronized data visualization
  • Safety-critical systems — geofence enforcement, e-stop integration, velocity warmup, watchdog processes

Backend Engineering

  • FastAPI, Celery, Redis — distributed APIs, async task queues, message brokering
  • REST API design and microservices architecture
  • PostgreSQL — schema design, normalisation, query optimisation
  • Streamlit — data dashboards and ML-facing UIs
  • Linux systems administration — Ubuntu, Arch, Fedora, Void (scripting, service management, performance tuning)
  • Networking — switches, structured cabling, 802.1x, IP configuration, system integration

Cloud & Infrastructure

  • AWS EC2 — instance provisioning, service deployment, remote access configuration
  • Azure AI Services — integration into backend processing and production ML workflows
  • On-premise GPU server setup and maintenance
  • MinIO — S3-compatible object storage for self-hosted deployments
  • Keycloak — identity and access management, OAuth2/OIDC integration
  • Infisical — secret management for multi-container deployments

Programming

  • Python — primary language for backend systems, ML tooling, and automation
  • Rust — systems-level desktop application development (Tauri)
  • TypeScript / React — desktop editor frontend, 3D viewport (React Three Fiber)
  • C / C++ — performance-critical and systems-level programming
  • C# — application development (Unity, XR platforms)
  • Lua — production reverse proxy scripting (OpenResty) and rapid prototyping
  • SQL — schema design, normalisation, query optimisation

Experience

Researcher — AI Optimisation (Postdoctoral)

Kingston University, London | 2024 – Present

RAIDO — EU Horizon Europe (Grant No. 101135800)

  • Designed and deployed a synthetic data generation pipeline using StyleGAN2, Stable Diffusion 1.5 with LoRA, and TimeGAN; served models via FastAPI + Celery + Redis + Flower in a Docker Compose stack on GPU and CPU
  • Built a Unity-based parameterised image capture tool using 3D digital twins, with configurable camera angles, sensor types (depth, IR, night vision), and environmental conditions (rain, fog, lighting)
  • Generative AI data augmentation improved image classification accuracy by up to 22.5 percentage points; subsequent model distillation reduced energy consumption by up to 75.8% and parameter count by 66% while retaining performance gains
  • Time-series MAE reduced by 52% through TimeGAN-generated synthetic data; distilled model size reduced from 2.28 MB to 0.77 MB
  • Deployed and integrated partner model distillation pipelines (PatchTST, MobileNetV3→MobileNetV2) into the same containerised stack
  • Integrated continuous learning components for image and time-series domains; deployed on AWS EC2, Azure, and on-premise GPU infrastructure

Researcher (Immersive and Intelligent Systems)

Kingston University, London | 2021 – 2024

Designed, deployed, and maintained ML systems across three funded projects spanning edge inference, generative AI, model distillation, and XR integration. Configured and maintained on-premise Linux servers and GPU workstations. Published 13 peer-reviewed papers across IEEE, Image and Vision Computing, and Journal of Systems Architecture. Initial development of the RAIDO data generation and distillation pipeline (continued under postdoctoral appointment).

TALON — EU Horizon Europe (Grant No. 101070181)

  • Developed a VR operator training application in Unity targeting Meta Quest 3 as part of the KU team within a multi-partner EU project
  • Built an edge inference backend running YOLOv5 with lifelong learning mechanisms; reduced AR detection transmission latency from 1 s to 200 ms (5×) to a HoloLens 2 client for real-time AR maintenance
  • Integrated ML backend services with XR frontends across a distributed research team on constrained edge hardware

Leonardo — University Contract

  • Developed a pre-processing pipeline for drone-captured imagery using ESRGAN super resolution to improve input quality for downstream small object detection in distant scenes
  • Implemented hyperparameter search for fine-tuning Faster R-CNN, YOLOv3, and RetinaNet on upscaled imagery
  • Research contributed directly to PhD thesis work on scene analysis for smart devices and immersive technologies

Publications

  1. Is it worth the energy? An in-depth study on the energy efficiency of data augmentation strategies for finetuning-based low/few-shot object detection — Journal of Systems Architecture, 2025
  2. Enhancing 3D object detection in autonomous vehicles based on synthetic virtual environment analysis — Image and Vision Computing, 2025
  3. Enhancing manufacturing training through VR simulations — IEEE International Conference on Engineering, Technology, 2025
  4. A complete survey on contemporary methods, emerging paradigms and hybrid approaches for few-shot learning — arXiv preprint, 2024
  5. Toward green and human-like artificial intelligence: A complete survey on contemporary few-shot learning approaches — arXiv preprint, 2024
  6. Evaluating the energy efficiency of few-shot learning for object detection in industrial settings — IEEE Real-Time and Intelligent Edge Computing Workshop, 2024
  7. A closer look at data augmentation strategies for finetuning-based low/few-shot object detection — IEEE International Symposium on Industrial Embedded Systems, 2024
  8. Scene analysis for smart devices and immersive technologies — PhD Thesis, Kingston University, 2024
  9. A modular deep learning framework for scene understanding in augmented reality applications — IEEE Industry 4.0, AI & Communications, 2023
  10. Evaluation of environmental conditions on object detection using oriented bounding boxes for AR applications — IEEE DCOSS, 2023
  11. Super resolution for augmented reality applications — IEEE INFOCOM Workshops, 2022
  12. Object recognition for augmented reality applications — Azerbaijan Journal of High Performance Computing, 2021
  13. Growth-based 3D modelling using stem-voxels encoded in digital-DNA structures — SIGGRAPH Asia Posters, 2020

Projects

Self-Hosted Infrastructure Stack

  • Designed and operate a full self-hosted service stack: API (FastAPI), async workers (Celery + Redis), object storage (MinIO), PostgreSQL, Lua-scripted reverse proxy (OpenResty), identity management (Keycloak + OAuth2), secret management (Infisical), monitoring (Grafana), and remote access (Guacamole)
  • Automated TLS certificate provisioning and environment-driven configuration across 27 Docker Compose service definitions
  • Deployed GitHub Actions CI/CD pipeline for automated builds, testing, and GitHub Pages documentation

Unitree G1 Robotics

  • Built a full robotics orchestration platform (editor + simulation + real hardware deploy) for the Unitree G1 humanoid robot, structured as a Unity-style authoring environment with a scene graph, component inspector, and one-click Play/Deploy workflow
  • Implemented a real-time perception pipeline: RT-DETR person detection at 10 Hz with ByteTrack multi-object tracking, bearing and distance estimation, and live bounding box overlay in the editor viewport
  • Wrapped a pretrained ONNX locomotion policy executing at 50 Hz; integrated a FollowPerson behavior tree (py-trees) with P-controller for bearing and distance, lost-target recovery, and safety gate with geofence and e-stop
  • Built a HAL abstraction layer over CycloneDDS enabling the same component graph to run in MuJoCo simulation and on the real G1 hardware with a configuration swap; implemented containerised one-click deploy with staged velocity ramp and SSH-based e-stop
  • All sessions (sim and real) recorded to MCAP and visualised in Rerun with synchronized camera, detection, robot state, and BT tick streams; integrated XR teleoperation via Meta Quest 3

Education

PhD in Artificial Intelligence — Kingston University | Awarded 2024

MSc Games Development (Programming) — Kingston University | 2019 – 2020

BSc (Hons) Computer Science (Games Programming) — Kingston University | 2015 – 2019

Additional

  • VR/AR development: Unity, Unreal Engine, Microsoft HoloLens
  • Hardware: server and workstation build and configuration, GPU setup, structured cabling and network switches
  • 3D modelling: Blender (surface design, STL export for fabrication)