PhD research internship

A researcher describes an experiment. Can a model reconstruct it?

Bower is looking for a PhD candidate to spend 3 months full time on vision-language models at the bench: the systems that read what a researcher says and shows while an experiment is running, and turn it into a structured research record.

The models already transcribe and extract. What nobody has measured properly is how much bench context they actually recover, where they fail, and whether a researcher can tell the difference. That is the project.

Apply by emailNo closing date

Key details

Project: Vision-Language Models at the Bench: Records Researchers Can Trust
Duration: 3 months, full time
Location: Remote or on site, negotiable
Eligibility: Enrolled PhD candidate
Field: Multimodal ML, computer vision, HCI, or a computational lab science
Start date: Negotiable. Applications reviewed as they arrive
Funding: Arranged through your university placement scheme
Intellectual property: Retained by Bower. A confidentiality agreement applies

Apply to

careers@bowerlabs.ai

Who you would work with

Bower builds the research record as the work happens

Researchers generate knowledge. Organizations lose it, somewhere between the bench and the write-up.

Bower captures scientific work while it is being 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 speaks up when something looks wrong, while the researcher is still standing at the bench and can do something about it.

How much of the bench a model actually understands is the foundation the rest of the product sits on. A record a researcher cannot trust is worse than no record at all, which is why this work is worth 3 months of a researcher’s attention rather than a sprint.

Who you would report to

David Lyon, co-founder and CTO. You would work with him directly, every week.
Renaud Joannes-Boyau, co-founder and Chief Science Officer, on scientific direction.

The project

Vision-Language Models at the Bench: Records Researchers Can Trust

Speech gives you what the researcher said. Vision gives you the bench they said it at: an instrument readout, a handwritten label, a plate that does not look right. A vision-language model has to hold both at once, and no standard benchmark covers that combination in a working laboratory. Over 3 months you would establish how much bench context a VLM actually recovers, and close the gap between that number and what a researcher needs before they will rely on the record. The goal is not a better transcript. It is a model that notices a protocol has drifted and says so while the researcher can still correct it.

Build the ground truth

Assemble an evaluation set from real protocols: bench recordings and images, paired with the record a researcher would have written by hand. Nothing here is a public benchmark, so the set has to be designed as well as collected.

Find where bench context is lost

Measure which parameters a VLM misses, and under what conditions. Background noise, an accent the model has not heard, a handwritten label at an angle, a reagent name that sounds like three other reagent names. Error analysis is most of the work.

Test whether one model beats two

A researcher says one thing and shows another, and the two disagree more often than you would expect. Establish whether a vision-language model reading both together recovers values that speech or vision alone loses, and quantify by how much.

Close the loop in real time

Accuracy only matters if it reaches the researcher while they can still act on it. Turn the model’s output into calibrated confidence and proactive deviation flags: a skipped step, a parameter drifting out of range, a reagent that does not match the protocol, surfaced at the bench while the experiment is running rather than in the data months later. This is the part that reaches the product.

What you get out of it

Three months that count as research

Real data, real researchers

You will evaluate against actual laboratory work rather than a curated dataset, with access to the researchers who produced it.

A study you run end to end

Scoping, collection, evaluation, and a written result. Industrial research on an industrial timeline, which is a different discipline to a thesis chapter.

Production ML experience

Vision-language models running against live capture, with the constraints that come with shipping: latency, cost, and a user who notices every error.

Supervision from the CTO

You would work directly with Bower’s co-founder and CTO, weekly, with scientific direction from the Chief Science Officer. Your findings change what gets built next.

Who we are looking for

We care more about how you think than which lab you sit in

Essential

enrolled in a PhD, ideally in your second year or beyond
strong Python, and hands-on work with vision-language or other multimodal models
able to design and run an evaluation, not only train a model
available 3 months full time
your university supports a placement of this length

Highly desirable

experience evaluating VLMs, or with speech recognition and OCR
a background in a laboratory science, or a track record working alongside scientists
exposure to annotation design or human evaluation
published work in a relevant venue

Research outcomes

What you would leave behind

The deliverables are agreed with you in the first fortnight, and they are the same things you would want on your own record at the end of it.

an annotated evaluation set covering a defined range of protocols
a written report on extraction accuracy, with error analysis
a working prototype that scores confidence and flags deviations in real time
a handover the engineering team can build on

And if your model outperforms what we are running, we integrate it. That means watching researchers use your work in their own laboratories before the placement ends, rather than reading about it in a report six months later.

How to apply

Send us an email. That is the whole process.

There is no portal and no closing date. We review applications as they arrive, and we will tell you either way.

Include

a short note on what you work on, and why this project interests you
your CV
your supervisor’s name, and confirmation that your university supports a 3-month placement

Apply by email

Supervisors and placement offices are welcome to get in touch at careers@bowerlabs.ai before a student applies.