CharliPosner

My passion for technology growing up led me to study CSE at the University of Bristol.

I studied everything from computer architecture, electronics and distributed systems, to cybersecurity, robotics and CGI. My favourite subject - computer vision - inspired both my Toshiba and Master's research projects.

University taught me hard work, discipline, and how to sit with difficult problems that seem impossible until they aren't. I graduated with First Class Honours, averaging 80%.

I spent a semester at Toshiba's Bristol research lab capturing a dataset of people walking using two calibrated Azure Kinect sensors and interleaving their captures to 60fps.

It was my first taste of real research: a vague problem with no guaranteed outcome. I ran the whole pipeline, from planning the experiment and capturing the data to training an LSTM.

The dataset is public on Zenodo, and our paper is published in Elsevier Data in Brief.

My Master's thesis asked whether a small AI camera, the OAK-D, could guide patients through physical therapy at home, rating each exercise out of 3 in real time, with no patient data leaving the home.

I turned pose data into images and retrained an image network (ResNet-50) to grade the exercises, then ran it live on the OAK-D alongside the pose tracking, at 20fps.

It was awarded Best Final Year Project, cemented my love of pose estimation, and inspired Human Sequencer.

I wanted to keep working in machine learning, so my first full-time role was as a Junior Software Engineer in the R&D team for KDB.AI, a vector database built on kdb+.

Before LLMs were widely known outside machine learning, I was building with them: integrations for LlamaIndex, LangChain and ChatGPT, and running models locally with llama.cpp.

I also researched text search algorithms (BM25, Hyperscan), and built an image search demo, using a ResNet trained on an MRI brain scan dataset.

Travelling and moving to Australia were always my dream. So less than a year into KX, I quit, backpacked solo through Southeast Asia for 10 months, then moved to Melbourne.

Along the way, I volunteered teaching English and computer skills at NDCCO in Siem Reap, got my advanced scuba certification, learned meditation and surfing, and tried Muay Thai, which I still train.

This experience significantly boosted my self-confidence and independence while providing clarity about my future goals; in particular, I knew I wanted to pursue a career in AI.

I joined Stile, a science education company, on a one-month contract to prove AI could scan students' handwritten worksheets into the platform. It worked: it became a full-time role, and eight months later I demoed the feature at NSTA in Minneapolis.

As a Builder in Stile's AI Labs, I work across the stack: AWS workstations for coding agents, pipelines that store and review LLM chat completions for fine-tuning, image editing with off-the-shelf models, and an iOS scanning app built on Apple's Vision SDK.

Applying agentic AI and harness engineering well lets me ship high-quality work fast, in almost any domain.

MLAI is a machine learning and AI community in Melbourne, now thousands strong. After going to their events when I first arrived, I started volunteering, and co-organised two large events: How To Make Your First Million and the AI Startup Pitch Competition.

At the pitch competition I was one of four judges, alongside three venture capitalists, there to give an AI engineer's view. I also went with MLAI to the World AI Conference in Shanghai, and shared my takeaways at their How Far Ahead is China? event.

What started as an idea between friends, to build a life-size music sequencer, became a fully interactive dancefloor installation at Lost Paradise, a music festival with 15,000+ attendees. Two AI cameras track people's movements, and the tiles underneath them light up and change the music playing.

I built its computer vision: real-time pose tracking on both OAK-D cameras, mapping people's skeletons to the floor tiles. I vibe coded it in two weeks, and it ran unattended for days at the festival.

Through MLAI, I met the creator of The Shape of Thoughts, a 12-day interactive exhibition in Melbourne centred around psychology and art. A friend and I turned one of her sketches into Confessions: a whirlpool of online confessions covering the walls of a room.

Each confession's colour shows the kind of secret it reveals, and its size shows how common the confession is in the dataset. I wrote about the data science and TouchDesigner network in this blog post.

It was my first time showcasing artwork in an exhibition, and working with a designer to bring her vision to life.

At Stile, I was tasked with exploring if AI could give our character illustrations new facial expressions. We only had smiling examples, too few to train a LoRA, so I turned to the research literature.

During my experiments, three principles emerged that allowed us to make style-faithful image edits: being less specific in prompts, describing movements rather than appearances, and running each feature edit in parallel from the original, then recombining.

I gave it as my first conference talk, and wrote up the method in this blog post, to show people an application of image generation they hadn't seen before.