Founding Applied AI Engineer

Tacit knowledge is the context gap in the lab. Build the visual perception layer that catches it.

Bower is hiring its founding applied AI engineer to own the perception stack: the video pipeline behind continuous capture, and the vision-language models that turn it into context an agent can act on.

The model is the part everyone talks about. The hard part is everything around it: video arriving faster than anyone can label it, storage that has to stay affordable for years, and a VLM that has to understand a laboratory well enough that an agent can be trusted to speak up about what it sees, in a real-time, low-latency conversation."

Apply by emailNo closing date

Key details

Role: Founding Applied AI Engineer, perception and video
Type: Full time, permanent
Location: Hybrid. South East Queensland or Northern New South Wales
In person: Regular days together on the Gold Coast/Byron Bay
Reports to: David Lyon, co-founder and CTO
Work rights: Australian work rights required
Compensation: Competitive package and meaningful founding equity
Start date: As soon as you are able. Applications reviewed as they arrive.

Apply to

careers@bowerlabs.ai

Why this company

Researchers generate knowledge. Organizations lose it.

Somewhere between the bench and the write-up, most of what a scientist knows about an experiment stops being recoverable. Bower is built to close that gap while the work is still happening.

Bower captures scientific work as it is done, through voice, images, and video, and turns it into a structured record the whole team can search. A researcher talks through a protocol without stopping to type. Bower transcribes it, extracts the parameters, links them to the experiment, and flags what looks wrong while the researcher is still standing at the bench and can do something about it.

Voice got that started. Vision is what makes it complete, because the bench holds things a researcher would never think to narrate: an instrument readout, a handwritten label, a plate that does not look right, a step performed differently to the way it is written down. Capturing that is a perception problem, and right now it is the biggest piece of the product without an owner.

Who you would work with

David Lyon, co-founder and CTO. Previously a senior engineering leader at Atlassian. He would be your manager, and your daily collaborator.
Renaud Joannes-Boyau, co-founder and Chief Science Officer. A geochemist and paleontologist who has run research laboratories for over 15 years, and the reason the product knows what a bench actually looks like.
Michelle MacRae , co-founder and CEO. Principal AI Product Lead at Microsoft for generative video.
Smart Glasses and Mobile Engineer, The in-house engineer building the smart glasses experience and the cross-platform mobile app.

What you would own

One stack, from the sensor to the agent

This is a founding role, so the boundary is drawn around a problem rather than a layer, with high autonomy to make fast decisions and deliver. Everything between the client side input channels and an agent that can provide meaningful context: the product features, the data processing pipeline, the storage, the models, the evaluation, and the cost of running all of it. The four pieces below are the shape of the first year, roughly in the order the difficulty sits.

Video at scale, and what it costs to keep

Continuous capture produces far more video than anyone will ever watch, and the bill arrives whether it gets watched or not. You would own ingest, storage tiering, and the indexing that makes one stretch of a recording findable months after nobody remembered taking it. Get this wrong and the result is not a slow product, it is a company that cannot afford its own data.

Context a model infers, and an agent can act on

This is the modeling problem the product sits on. A VLM has to look at a bench and work out what is actually happening: which step of the protocol, which reagent, which instrument reading, whether the vessel on the stir plate matches what the researcher described a minute ago. Every agent downstream inherits that context, and inherits it being wrong.

Accuracy per dollar, until watching everything is affordable

Frontier VLMs can be both accurate and expensive, and capture runs all day. Quantization, distillation, frame selection, caching, routing a question to a small model and escalating only when it matters. The target is not a benchmark score, it is a cost curve that lets Bower look at all of the work rather than a sample of it.

Perception at the bench, in real time

Some of this has to run on the glasses, on hardware with a battery and a thermal budget, against a researcher whose hands are full. You would decide what runs on device, what goes to the cloud, and what the latency budget is for someone who needs to be told now rather than tonight.

Why this role

Early enough that the decisions are still yours to make

The stack has no owner yet

Perception is unclaimed. You would set the architecture, choose the models, and decide what gets built and what gets bought. Founding engineer means those calls are genuinely yours to make, and genuinely yours to live with.

Hardware nobody has settled

Smart glasses in a working laboratory are early enough that the interaction patterns are still open questions. Some of what you would work on has no published answer, which is the good kind of hard.

Users you can stand next to

Real researchers, in real laboratories, at the bench. You would watch your model be wrong in front of the person it was wrong about, which is the fastest feedback loop this work has.

Compressing the time to benefit

Between a result in a laboratory and anything that reaches a person the needs it, there is a handoff that loses most of what made it work. Bower exists to shorten that. Bring discoveries to humanity at an accelerated pace.

Who we are looking for

We care what you have shipped, not how long you have been shipping

There is no number of years on this list, deliberately. What matters is whether you have taken a vision system to production, watched it fail in ways nobody predicted, and fixed it. If that describes you after four years, apply. If it took you fifteen, apply.

Essential

deep hands-on work with vision models or VLMs in production, not in a notebook
you have owned a video or large media pipeline end to end, including what it cost to run
systems thinker, building AI automations not only around your coding but all your workflows and life
you optimize models as a matter of course: quantization, distillation, batching, frame selection
you ship, which includes evaluation, deployment, monitoring, and the parts that keep a model honest once real people depend on it
Australian work rights, and based in South East Queensland or Northern New South Wales, or willing to relocate

Highly desirable

on-device or edge inference, especially on constrained or wearable hardware
building agentic systems on top of a perception layer
video understanding, temporal models, or multimodal retrieval
a background in a laboratory science, or a track record working alongside scientists
you have been an early engineer somewhere before and know what it actually costs

How we work

We work the way a lab works, because one of us ran them for 15 years

Bower was started by a geochemist, a product leader, a go-to-market operator, and an engineer, and the way the company runs still borrows from the bench: form a hypothesis, run the smallest experiment that would disprove it, and change your mind when the result says to. These four are the ones you would feel in an ordinary week.

Follow the evidence

We measure, then decide. A technical position with no number behind it is a preference, and we are expected to say so when that is what it is.

Begin with the scientist

Every choice gets judged on whether it helps a researcher at a bench. Not on whether it is interesting, and not on whether it demos well.

Accuracy is sacred

A record a researcher cannot trust is worse than no record at all. We show confidence, we surface what the model was unsure about, and we never present output as verified truth.

Stay humble and exothermic

Energy shared multiplies. We are grounded in getting it right rather than getting credit, and the work is allowed to be fun while the science stays serious.

Before you apply

What this job is not

A founding role is a specific thing, and it suits a specific person. Here is the part most postings leave out, so you can rule yourself out now rather than in month three.

not a research position. There is a research track at Bower, the PhD internship, and this is not it. Most days you would be writing production code
not a role with a team under it on day one. That may change, and if it does the team would be yours to build, but come expecting to build the thing yourself first
not a clean dataset. No public benchmark describes a working laboratory well enough, so part of the job is deciding what good means and building the evaluation that measures it
not settled. Priorities move, and some weeks the right answer is different hardware, a vendor, or deleting what you built last month
not for someone who needs the problem handed over already scoped

If none of that put you off, the upside is the same fact read the other way.

How to apply

Send us an email. That is the whole application.

There is no portal, no closing date, and no form that loses your work when the session expires. We review applications as they arrive.

Include

a short note on what you have built, what your role was in that team, how it relates to the skills Bower needs
your CV, a GitHub, or whatever actually represents the work
one system you took to production, and what you would do differently now

Apply by email

Not sure whether your background fits? Write anyway. The list above is what we are looking for, not a scoring rubric.