About
Main bio
i spend my days on the part of generative AI that doesn't show in the final image: datasets, captions, training runs, and the checks that stop a character drifting from one output to the next. the people using what i build are designers and marketers with deadlines. my job is to make the system work so they don't have to think about it.
i've been the only engineer on a project, and i've worked with creative teams who know the brand better than i do. i've trained models on local and cloud GPUs, built the ComfyUI workflows that turn checkpoints into output, and shipped the app that hands it over. before AI, i spent years building and measuring websites, which is why i care what happens after something ships.
currently i'm at Bernadette (VCCP), building a generative AI character pipeline for a brand campaign. i can't name the client, but the full case study is on a password-protected page and i'm happy to give access. before that, i built AI automation for UK agencies through Seenovate.
it's me
HOW I WORK
start with the data, not the model
i'd rather spend the first days on what goes into a model than on its settings: which images, how they're captioned, what it should learn and what it shouldn't. a character that looks almost right usually started with a dataset that was almost right.
data first, train later.
no perfect world here!
build for the setup you have
a campaign deadline, a client i can't name, and a GPU budget. i've trained on a local RTX 4090 and on Azure GPU VMs, and the work has to fit the hardware, the schedule and what i'm allowed to show.
built with the people who use it
designers and marketers know the brand better than i do. i build the rubrics, workflows and checks around what they tell me good looks like, and the test is whether they can run it without me.
we, not me.
checks build trust.
make the checks visible
the most useful parts of the pipeline never show in the final image: caption validation, a vision QA agent scoring each output against a per-character rubric, and workflows versioned so a run can be reproduced. they catch drift before a person reviews anything.
EXPERIENCE
Timeline
Bernadette (VCCP)
building a generative AI character pipeline for a brand campaign (client confidential). FLUX LoRA training, ComfyUI workflows for the design team, vision QA on every output, and a self-serve app so the marketing team can produce approved content without an engineer.
MAY 2026 - OCT 2026
Seenovate
AI automation and product engineering for UK agencies and startups. white-label n8n workflows for reporting, invoice chasing and onboarding, the CardioCare dashboard (FastAPI, React), and a multi-agent harness where one agent builds and another reviews through pull requests.
Feb 2025 - Current
Lierex
senior web builds for client brands: design systems, WooCommerce and WordPress sites, A/B tests on landing pages and Core Web Vitals work. this is where i learned to check what happens after launch.
SEP 2023 - FEB 2025
WebRGB
50+ WordPress and WooCommerce projects, with GA4 used to find where people dropped off and fix it. this is where i started measuring results instead of guessing.
DEC 2019 - SEP 2022