Machine Learning Engineer · London
Lolézio
Viora Marquet
I build machine learning systems and make them run fast. Computer vision and neuro-symbolic research on one side, GPU kernels and embedded C++ on the other.
First Class MEng from Imperial College London. Software Development Engineer at Amazon. Looking towards machine learning and computer vision roles.
Selected work
Everything, most recent and most substantial first.
- October 2025 – June 2026 Research
Neuro-Symbolic Learning under Multi-Modal Data
Master's thesis: teaching models to be right for the right reasons
MEng thesis on neuro-symbolic AI: getting models to reason over the correct underlying concepts instead of exploiting shortcuts, even when one concept appears in many visual forms. New state of the art across several established neuro-symbolic frameworks; the paper is under review at a leading AI conference.
- SOTA across multiple NSAI frameworks
- OOD gains in and out of distribution
- February 2026 – March 2026 Coursework
AutoFuser
Automatic Triton kernel fusion for PyTorch models
A team project for Imperial's Advanced Deep Learning Systems course: a compiler pipeline that finds chains of operations in a PyTorch model, generates and auto-tunes a fused Triton kernel for each, and rewrites the model to use them. I proposed the design and built the code generation, tiling, autotuning and graph rewriting stages.
- 3.7× faster inference on the Galerkin Transformer, from fusion alone
- 1.71× on a transformer with MXINT8, ahead of torch.compile's 1.61×
- September 2026 – Present Live
Agentic Coding Harnesses at Prime Video
Agents that write and review production code for a page millions of customers use
Custom agentic coding harnesses built around my team's development process on the Prime Video Detail Page: agents write and review code changes end to end, running as long-lived sessions on a remote cloud desktop, so my manual work is steering and final sign-off.
- millions customers on the page I ship to
- write + review code changes handled end to end
- July 2024 – Present Live
Cryptocurrency Trading Bot
A transformer trading live capital, 24/7, since 2024
A systematic trading bot built solo: a transformer that predicts crypto market movements and executes live trades through the Binance API from an EC2 instance, with a full backtesting and deployment pipeline behind it. Averaged 30% returns.
- ~30% average returns
- 24/7 live in production since 2024
- April 2025 – September 2025 Shipped
Amazon Fuse Internal Platform
A serverless internal tool, from requirements to production
An internal full-stack web application built end to end during a six-month internship on Amazon Fuse: a React frontend over a native AWS backend, with gated APIs letting authorised users query, edit and delete production database entries. 3,000+ daily requests at sub-300 ms latency, and still in daily use a year after handover.
- 12k requests at peak, sub-300 ms
- 150 daily users
- February 2025 – March 2025 Shipped
STM32 Synthesiser, running DOOM
An embedded synthesiser on an STM32, and the 3D game engine I wrote to run on it
Team coursework in embedded systems: a real-time FreeRTOS synthesiser on an STM32 microcontroller. I built the user interface and system state machine, and wrote a DOOM-style 3D game engine from scratch because the smallest existing port of DOOM is 1.4 MB.
- 45 fps game frame rate, with a reduced render distance
- 48 ms worst-case full-screen refresh, about 18 ms typically
- June 2024 – August 2024 Shipped
Offshore Wind RAG System
Retrieval-augmented generation over thousands of bid documents
A cloud-hosted RAG application built at Equinor for the offshore wind bid team: NLP retrieval surfaces the relevant documents, an LLM generates a grounded answer citing its sources, and a custom scraper builds the document base. Hours of manual search became a few-second query.
- 1000s of documents searchable
- seconds replacing hours of manual search
- May 2024 – June 2024 Shipped
Game of Life on FPGA
Conway's Game of Life in custom hardware, with a hand-gesture interface for drawing the grid
Second-year team project at Imperial: Conway's Game of Life accelerated on a PYNQ-Z1 FPGA board, computing a full 1280×720 generation in 722 clock cycles. I wrote the CPU benchmarks that sized the problem, the next-state logic, the grid initialisation and the hand-gesture interface.
- 722 clock cycles per 1280×720 generation
- ~200k generations per second in theory, at about 150 MHz
- January 2024 – March 2024 Coursework
C90-to-RISC-V Compiler
A working C compiler, written in C++
A compiler in C++ that translates pre-processed C90 source into RISC-V assembly: lexer, parser, internal instruction representation and code generation, validated against an extensive automated test suite. Core implemented in 36 hours.
- 36 h for the core implementation
- C90 source language
- October 2023 – December 2023 Coursework
Pipelined RISC-V CPU
A 32-bit CPU in SystemVerilog, built three times: single-cycle, pipelined, then with a cache
Second-year team coursework at Imperial: a RISC-V CPU in SystemVerilog, first single-cycle, then pipelined with hazard handling, then with a data cache. I built the program counter, the pipeline registers, the F1 lights program and most of the build and test tooling.
- 4 people on the team
- 5-stage pipeline with hazard handling
- January 2025 – February 2025 Shipped
PiTrainer
A connected gym tracker: Raspberry Pi sensors, a cloud backend and a mobile app
A three-person IoT project: a Raspberry Pi with motion sensors counts reps on gym machines, a Flask backend on AWS scores each rep with a machine learning model, and a mobile app controls the workout and shows the analysis. I built the backend, the rep-quality model and the mobile app.
- 3 workout states synchronised between app, cloud and Pi
- 96 statistical features per rep for the quality model
- January 2023 – June 2024 Shipped
Project ATLAS
Leading 15 engineers to build an autonomous drone
Led a team of 15 engineers at the Imperial College Drone Society building an autonomous drone that crosses an obstacle course using machine and deep learning, custom electronics and an efficient in/out system. Raised and managed over £1,000.
- 15 engineers led
- £1,000+ raised and managed
- February 2024 – March 2024 Shipped
2-Player Flight Simulator
FPGA joysticks, an AWS server, and two planes in one world
A two-player racing game where each player flies using a DE10-Lite FPGA as a physical controller. The FPGAs connect wirelessly to an AWS EC2 server that processes both players asynchronously and renders them into a shared Unreal Engine world.
- 200 ms end-to-end input latency
- 2 players in a shared world
- November 2023 – January 2026 Shipped
Maisha Design Platform
Generative AI and a full-stack platform for an interior design firm
Two years building the software that runs an interior design firm: conditional GANs applied to the creative process, a platform centralising supplier furniture catalogues, invoice generation, and scrapers with LLM-enhanced category matching. Cut project turnaround by 30%.
- 30% faster project turnaround
- 2 yr engagement, in production
- May 2023 – June 2023 Shipped
Fyrryx Rover
A remote-controlled exploration rover; I built its magnetic field sensor and sensing firmware
First-year team project at Imperial: a remote-controlled rover that identifies 'aliens' by their radio, infrared and magnetic signatures. I designed the magnetic field detector, wrote the firmware for the sensing board, and moved the sensors onto a second microcontroller linked over I2C.
- 12 cm magnetic detection range, against a 5 cm spec
- 9.74 ms per field reading, inside a 10 ms budget
- August 2020 – February 2022 Archived
Search-and-Rescue Drone Vision
Fast obstacle detection for drone flight with YOLOv3-tiny, trained on my own dataset
An independent research project on making drone collision avoidance faster: a YOLOv3-tiny detector trained on a dataset I built and annotated, and a lightweight steering algorithm driven by its detections. Overall winner of the Vienna International Science Fair 2021.
- Overall winner Vienna International Science Fair 2021
- 50,000+ training iterations on a dataset I annotated
- 2023 Archived
Autonomous Greenhouse
Closed-loop environmental control to increase plant yield
An autonomous greenhouse that raises plant yield by sensing and controlling its own temperature, humidity, light and watering in a closed control loop, with the control firmware in C++.
- closed-loop environmental control
Experience
-
Software Development Engineer at Amazon
On the Prime Video Detail Page team, shipping customer-facing features to production for millions of customers, with AI coding agents as my default workflow. Joined full-time on a return offer from my Amazon Fuse internship.
-
Software Development Engineer Intern at Amazon
Built an internal full-stack web application end to end, from user requirements to production, on a native AWS stack.
-
Software Engineering & Data Science Intern at Equinor
Built a cloud-hosted RAG system that turned hours of manual document search into a few-second query, for a team bidding on offshore wind farms.
-
Software & Machine Learning Engineer at Maisha Design
Two-year engagement applying machine learning to the creative workflow of an interior design firm, and building the software around it.
Skills
-
Machine learning
Research through to production
-
Low-level & performance
Where the cycles actually go
-
Hardware
Digital design and silicon-adjacent work
-
Product & cloud
Shipping things people use
-
LLMs & agents
The default way I work
-
Also
Picked up along the way
Education
-
Imperial College London
First Class Honours- Core engineering across mathematics, AI/ML, hardware, software, networks and databases.
- Master's thesis on neuro-symbolic learning under multi-modal data, reaching a new state of the art. Paper under review at a leading AI conference.
-
Vienna International School
43 / 45- Higher Level: Mathematics Analysis & Approaches (7/7), Physics (7/7), Chemistry (7/7).
- Standard Level: Economics (7/7), Spanish (7/7), English Language & Literature (6/7).
- Extended Essay in Physics. SAT 1560 (2021).
Looking for ML and computer vision work
Most interested in roles where research and systems engineering meet: vision, neuro-symbolic reasoning, or making models run fast on real hardware. If that's what you're building, get in touch.