Senior AI/ML Engineer

About the role

You will own damage detection across SyncTech's platform. Somebody captures a damaged property, either with a 3D scanner or a phone. From that, we build a model of the space, find every instance of damage, produce a report an assessor can read, measure how much of each surface is affected, turn that into a priced scope of work, and learn from the assessor's corrections.

Finding the damage is the heart of it. Not "this photo shows water damage", but each individual instance, with what it is, what material it is on, which surface, which room, and where in three dimensions. A single photo usually shows several distinct damages. One patch of wall can be two kinds of damage at once, so naming is multi-label rather than pick-one. The same crack turns up in fourteen frames and is still one crack. And the imagery is real claim photography: badly lit, motion blurred, shot at awkward angles by people who are not photographers.

You will fine-tune vision language models on this imagery, decide the framing rather than inherit it, get from per-image predictions to a claim-level result, set the labelling standard the models are trained on, and build the evaluation that tells us whether any of it is working. The programme runs across all seven steps and the team is small, so the boundary will move. Depending on where you are strongest, that could grow into measurement, which means turning a located damage into a defensible quantity and is the piece the entire scope of work is waiting on. It could also mean feeding assessor corrections back as training signal, or the pipeline that runs a whole claim end to end.

Our work goes into live claims, where a model's output feeds decisions about real money on real buildings. Our clients hold us to agreed accuracy standards, so it is never enough for something to look like it works. We have to be able to show that it does. Insurance is a heavily regulated industry and the imagery we work with is the inside of somebody's home. Privacy obligations shape what data we are allowed to use, where it is allowed to go, and how we build.

Key responsibilities

• Fine-tune vision language models to find each instance of damage in an image and name it, with the material, the surface, and a confidence. Choose the framing: direct detection, or a detector proposing regions the model describes.

• Establish whether coordinate precision is good enough for fine detail like hairline cracking before we build on the assumption that it is.

• Merge repeat sightings of the same damage across frames, pin each one to a place in the building, resolve which room it is in when rooms run into each other, and score confidence so a human only checks what is worth checking.

• Set a labelling standard that matches the problem, which means per damage, multiple types allowed, room recorded. Work out what the existing label set is worth and what it would cost to redo.

• Design around the pipeline that finds and removes personal information before training and again before a report goes out. Know what may go into a training set, what may leave our environment, what may never touch an external service, and when to stop and ask.

• Define what accuracy means per damage type, what counts as a match, how a partly correct multi-type answer scores, and whether we measure per damage or per image. Build the harness that answers those questions repeatedly and cheaply.

• Notice the thing nobody has looked at, decide whether it is worth a week, and find out. Time-box it, write it up, and report it whichever way it turns out.

About you

• A bachelor's degree in engineering, computer science, software engineering, data science, mathematics or a related technical field, or equivalent practical experience.

• Substantial hands-on experience fine-tuning vision or vision language models on domain-specific imagery, with results you can talk through in detail.

• Detection and segmentation depth beyond image-level classification. Multi-label, multi-instance problems where the unit is the object rather than the picture, and where two labels can be correct for the same region.

• Comfort with the modern fine-tuning stack. Parameter-efficient methods, quantised training, and tooling in the family of Unsloth, LLaMA Factory or TRL. Understanding of what these do and when they break.

• Judgement about when to unfreeze a vision tower, when to train language layers only, how much labelled data a result actually needs, and when the answer is better labels rather than a bigger model.

• Genuine depth in evaluation. Ability to explain why a benchmark comparison did or did not support its conclusion, and experience arguing that case with people who did not want to hear it.

• Data-centric instincts. Experience looking at training data, finding what was wrong with it, and fixing that before reaching for a new architecture.

• Working understanding of privacy obligations around personal information. Thinking about what is actually visible in an image before it goes into a training set, and about where data ends up when reaching for an external tool.

• Engineering that holds up. Python, version control, reproducible training runs, experiment tracking, Docker and containerised workflows, GPU infrastructure. Knowledge of roughly what experiments cost and factoring that in.

• Hands-on experience with AWS ML and AI services such as SageMaker and Bedrock, or close equivalents. Experience training, tuning, deploying or serving models on managed cloud infrastructure.

• Good understanding of ETL and data pipelines for ML, covering ingestion, transformation, versioning, and what breaks. Practical experience with AWS infrastructure in the family of S3, Lambda, Glue or Step Functions, and judgement to recognise when the pipeline is the problem rather than the model.

• Track record of running own investigations to a conclusion, including ones that ended your own idea. Ability to articulate what you set out to test, how long you gave it, what you found, and what you did when the evidence went against what you wanted.

• Clear communication. A result nobody else can act on has not been delivered.

About us

Insurance decisions today are often made on fragmented information with little visual context. SyncTech changes that. Our platform combines visual-first workflows, digital twins, AI-driven damage detection, and remote inspection technology to help insurers, builders, and assessors make faster, clearer, and more defensible decisions.

We are past early product–market fit and already working with major enterprise insurers in Australia and Europe. Now we are entering the next phase of our journey: scaling adoption, expanding product capabilities, and building the category leader in visual property intelligence. We are building a company for the long term — and looking for people who want to help shape it.