Rosie is a Staffy and shar pei cross from Sydney. For about a year, her owner Paul Conyngham was told that the lumps spreading along her hind legs were just a rash. When the diagnosis finally came, it was mast cell skin cancer, Rosie was five, and the disease was incurable. Surgeries, chemotherapy and immunotherapy slowed it down but did not stop it. By mid-2025 her vets estimated she had between one and six months left.

Conyngham is not a biologist. He is an AI consultant and entrepreneur who, by his own account, has spent 17 years in machine learning. Professor Pall Thordarson, director of the UNSW RNA Institute, believes he is the first non-scientist in the world to design and deliver a fully personalised cancer vaccine of this kind. Here is how he did it, and what is still unknown.

Step one: ChatGPT as the planner

Conyngham started with ChatGPT. According to UNSW's first report on the case, from June 2025, he used it to understand Rosie's treatment options and to draw up a plan. The first step was sequencing: reading the DNA of both Rosie's healthy tissue and her tumour. On ChatGPT's recommendation, he contacted the Ramaciotti Centre for Genomics at UNSW, led by Associate Professor Martin Smith, and paid about $3,000 for the work.

The centre sent back about 320 gigabytes of raw data. ChatGPT told him what to do with it, UNSW wrote, but he still had to work through the files and run the analysis programs himself, with no training in bioinformatics.

Step two: AlphaFold, and a dead end

From published research, Conyngham knew that a protein called c-KIT plays a key role in mast cell tumours in dogs. He used AlphaFold, Google DeepMind's software for predicting protein structures, to model Rosie's version of c-KIT and found that it looked abnormal. He then found an existing drug, used in the United States for a different human cancer, that targets a similar protein.

That route ended when the drug's manufacturer declined to supply it for compassionate use in a dog, according to the Ramaciotti Centre's 2024–2025 biennial report. UNSW structural biologist Dr Kate Michie warned that AlphaFold can get things wrong and its models need lab confirmation; the model of Rosie's c-KIT had a low confidence score of 54.55.

Step three: a vaccine written for one dog

Conyngham changed approach. By comparing the tumour's DNA with the RNA it was producing, he looked for mutations the cancer was actually using, and found seven strong candidates for neoantigens: markers found only on tumour cells, which the immune system can be taught to attack. The Scientist reports that ChatGPT helped him identify the neoantigens and the vaccine sequence.

The Ramaciotti report adds a detail most coverage missed: the final vaccine construct was designed using Grok, the assistant from xAI. In an interview with AFP, Conyngham named ChatGPT, Gemini and Grok as the chatbots he worked with at length.

AI did not make the medicine. The UNSW RNA Institute agreed to produce the vaccine after an extensive review; according to The Scientist, its team made the mRNA from a DNA template and packaged it in lipid nanoparticles. Conyngham spent hours a day for three months on the ethics application, which he says was more work than designing the vaccine. The vaccine was given by the lab of Professor Rachel Allavena at the University of Queensland's School of Veterinary Science, which had ethics approval to test immunotherapy in dogs. Rosie received her first injection in December 2025, with boosters planned for early 2026.

What happened next

Alongside the vaccine, Rosie was given a checkpoint inhibitor, a separate immunotherapy drug that releases the brakes on the immune system. A few weeks later the tumours on her legs swelled, then began to shrink. In March 2026, UNSW described a partial response: the largest tumours shrank so much that her ankles were visible again, and she was running around the dog park. Some tumours did not respond at all. In his August launch post for his new company on Y Combinator's site, Conyngham wrote that several tumours shrank by 50–75%. That figure is his own and has not been published in a peer-reviewed paper.

The caveats

This is one dog, and the people involved say so. UNSW's own article lists what nobody knows yet: how long the response will last, why some tumours responded and others did not, and how much of the effect came from the vaccine rather than the checkpoint inhibitor. Asked by AFP what caused the tumours to shrink, Thordarson said the "short answer is we don't know for sure".

Martin Smith told the ABC that this was a citizen science project, not a clinical trial. Nick Semenkovich of the Medical College of Wisconsin, who was not involved, told AFP that no scientific details had been published beyond press releases and interviews, so it is unclear how much the AI helped.

Experts handled the parts that require expertise: sequencing, manufacturing, ethics approval and veterinary care. By Conyngham's estimate, the whole effort took a year and around $20,000.

The story has since become a business. Conyngham founded Gamgee, part of Y Combinator's Summer 2026 batch, which designs personalised mRNA vaccines for dogs. It says a second-generation vaccine is ready and that work on a clinician-led trial in 30 dogs is well underway. The most recent independent report we found, from the ABC in June 2026, said Rosie still had cancer and several months to live. We could not confirm her condition since then.

What this means if you use AI for a complex personal project

Use a general assistant to plan and learn. Conyngham started knowing nothing about DNA or cancer. ChatGPT helped him understand the options, work out what data he needed and find a lab that could produce it. General assistants such as ChatGPT and Grok are good at turning an unfamiliar field into a list of concrete next steps.

Use specialised tools for specialised work. A chatbot can explain how proteins fold; AlphaFold, a dedicated scientific tool, actually predicts the structure. The same pattern holds in other fields: a general assistant can tell you which specialised tool you need, but it does not replace it.

Treat AI output as a draft for experts to check. In Rosie's case, scientists reviewed the design, made the vaccine and gave it under an ethics approval. In any high-stakes area, and above all in health, AI can speed up your research, but the decisions belong with qualified professionals.