The Quiet Disruptor - Inside an Interns AI Innovation

"If we can scale AI for better use, it can make the world a better place rather than just your daily personal life."

Notable Moments:

[01:27] Marina explains her AI project with indigenous communities in New Zealand

[03:26] Researching AI responses in abortion counseling

[07:43] Ten weeks in Namibia studying environmental science and conservation

[10:46] Using machine learning to study tree pod dynamics and herbivore consumption

[12:24] Expanding AI research into plant growth and climate variability

[13:25] Marina’s perspective on scaling AI to make the world better

In this bonus episode, Jody and Meghan talk with Redox intern Marina Frayre about her innovative work at the intersection of AI, healthcare, and environmental science. From empowering indigenous communities in New Zealand to studying conservation in Namibia, Marina shares how AI can both protect ecosystems and influence patient care. She emphasizes the importance of cultural respect, accuracy, and using technology to create meaningful global impact.

Resources

 www.redoxengine.com

Past Podcast Episodes 

https://redoxengine.com/solutions/platform-security

Have feedback or a topic suggestion? Submit it using this linked form.

Matt Mock  [e-mail] 

Meghan McLeod [e-mail]

 

 

Episode Transcript

(00:00):
Foreign.
Welcome to Shut the Back Door, brought to you by Redox.
Shut the Back Door is a healthcare security podcast dedicated
to keeping health data safe one episode at a time.
I'm your host, Jody Mayberry, and joining me,
as you would expect, is Megan McLeod. Hello, Megan.

(00:22):
Hi, Jody. Megan is always our friendly voice
from Redox. It just wouldn't be an episode of Shut the Back Door without
Megan. So I'm so glad you're here and I'm also glad at
our guests. We have a guest for this episode
which I want to point out is a special
episode. It's a bonus kind of mini episode

(00:44):
that we're trying out and we may do more of these in the future.
And our guest is Marina Freya
from. Well, she's an intern at Redox, but she's done a lot
more. She's currently an incoming senior at
Dartmouth College. I was about to say from Dartmouth, but she's also from
Redox. She's double majoring in computer science and

(01:06):
environmental science. Two things that seem like
opposites, but I'm sure Marina can tell us they actually go
hand in hand together. Welcome to the show, Marina. Hi,
Jodi. Thank you for having me. Yeah, it's great to have you here.
So we want to talk to you about your work in
AI. Can you talk about what you've done so far? Yeah,

(01:27):
I'd love to. So my work in AI has been more research
focused, but I would say equally as impactful. One of
my major projects has been working with two indigenous communities
in New Zealand, where I developed a series of
python scripts to fully automate environmental monitoring
workflows and ArcGIS. So the goal was to streamline the process

(01:51):
of multispectral satellite images and high
resolution drone data so that communities could track
environmental changes in near real time without needing to
rely on slow manual analysis. And these
workflows can be used to flag changes in, like the canopy
health, map erosion risk, and identify areas where

(02:12):
invasive species were spreading. And this is all very important
because when we all look at the ground, we obviously can't see below it. But
through these images we get a more full picture
view of this. And we were able to also track a very important
indigenous tree species and find out if it's still growing
and if its roots are healthy underground, which is important because

(02:34):
we don't know until the tree comes up that is up there. But now there
is a way through these satellite images to be able to track if that
tree is currently growing or healthy or if it's a concern
that has to be dealt with. And a key part of this
project was ensuring that work was respected and incorporated
indigenous knowledge and follow the indigenous data sovereignty

(02:57):
principles. This meant building systems that gave
the communities full control over how data was stored, shared
and used, while also making the output easy to
interpret without requiring advanced technical
training. It was a balance between, I would say cutting edge like geospatial
automation and. Deep cultural respect ensuring like

(03:19):
technology serve the community's priority rather than overriding
them. And another project I have worked on. More
in the healthcare space is where we
evaluated how large language models responded to
abortion related counseling questions. And this was super cool to
look into because we've seen that a lot of people have been turning

(03:41):
to either ChatGPT, Gemini or any LLM to
share their issues, especially health related.
So we wanted to see how these LLMs dealt with
questions that were more emotionally charged rather than there
being a yes or no factual answer. So
this work, yeah, was just not designated to test the

(04:04):
factual accuracy, but also whether the AI generated responses
were clear, fair, unbiased and emotionally appropriate
for the sensitive topic. We use a diverse set of prompts
covering different patient scenarios that people had shared from their
own personal experiences. And this was, we focus on the
US just because there's 50 states, but within the 50 states we know that

(04:26):
everyone has so many different laws and it's constantly
changing. So we also thought that would play a big role as well as
people have their own personal beliefs. And then from this
we examined the responses for signs of the
political bias, assess whether the information was presented in a
way that patients could actually easily understand it, especially

(04:48):
given the complexity of medical topics. I think we often find that people
share that they didn't fully understand what the doctors told them, but they just
kind of nod their heads and they're like, okay, like I'll just move on with
it. I can google this at home. But it was very important for us to
see whether or not the outputs that were being outputted initially
required people to google more or keep like talking to

(05:10):
the LLM or if it was generally like pretty
clear for people to understand. And we identified any gaps in
factual accuracy and noted variations also in tones that might be
unintentionally influencing a patient's decision making.
Especially with abortion, a lot of people are looking for
advice that is going to help them in the moment.

(05:32):
So we needed to find and make sure that these models weren't
giving certain types of tones where it might push a patient to do
one situation versus the other. Of course, like what's Right and
wrong depends on what you believe in, who you are. So that
was an issue we were dealing with. But the goal was just to highlight
where these models might inherently provide harmful or incomplete advice

(05:55):
and to propose concrete improvements both in model training
and in how AI systems are deployed in clinical or counseling
context so they can be safer and more trustworthy in higher stakes
healthcare settings. If that's what people choose to go with.
Yeah. So I think, Jodi, your question was answered pretty quickly on the
computer science and environmental science connection there. But

(06:17):
it's also really cool and valuable, you know, for us at
Redox to have your, also your experience
with healthcare, AI and stuff like that specifically since,
you know, like we are a healthcare data company. So you kind of have like
a, a great range, I think, of knowledge and experience
there to combine, because AI is in kind

(06:39):
of all sectors. So having like a
computer science background to be able to look at this from like a more technical
sense, but then also having the other kinds of backgrounds to be able to
take a look at this data and actually make something meaningful out of it, I
think that's just like a very valuable asset to have. Yeah.
I loved hearing about how AI is being used in

(07:01):
conservation with tree health. When I was a park ranger, you
would just go find the park ranger with the most experience and say, look at
this tree, Is it healthy? And he'd say yes or no. So this,
your model using AI is. I can see where it can
change conservation, it can speed up conservation, it can allow you to get
to problems long before you would have just doing the

(07:23):
eye test. Because sometimes when you see a tree's going bad, it's too
late. And with some of the other work you're doing, that can be true as
well in healthcare, maybe it can pick up problems before it's too
late. It's just fascinating. But I've also heard that
you've taken research abroad. Tell us about that experience.
Yeah. So this past fall I spent 10. Weeks in

(07:46):
Namibia on an immersive field based environmental
science program with Dartmouth. And it was 15 of us
going on this program. And it just combines scientific research with
direct engagement with stakeholders. So one week we might have
been on the field collecting measurements in the desert, and next we'd
be meeting with the government ministries, conservancy

(08:09):
leaders, local farmers, and discussing conservation policy
and land management challenges, which was super fun
and a very challenging program. Sleeping intense a lot of the time.
But I think that just added to the whole immersiveness
of this program. And while on there A project
I worked on that we called Pod Dynamics in the Namib

(08:31):
Desert is where we studied two key arid land
species, Vidyarbia in acacia, and assess how tree
size and sand density influence pod production and
herbivore browsing patterns. And this project was very important
because the farmers rely on these two trees because they produce
pods for the animals to eat. So they're basically service food

(08:54):
year round, and they both have different times of years where
they produce these pods. Therefore, there's year round availability.
And if there were to stop that availability,
then farmers would have to find an alternative way to feed their
cattle, which can come at a very expensive, especially if you have a lot of
it and being used to not having to pay for that. So that was super

(09:17):
cool. And so, yeah, so these pods are just very, very critical
source of nutrition for both livestock and wildlife during the dry
season especially. So understanding their availability
has real implications for food security and
conservation management. The project involved a
combination of detailed field work, measuring canopy

(09:39):
spread, trunk diameter and plot counts, paired with image
analysis and statistical modeling. We did have
access to pictures that were taken looking directly at the
tree and at the ground that would take every single day.
And we had these images for three years. We have
a data set of three years. We had to analyze this in three weeks, which

(10:01):
is a very fast phase project. But with this, we were able
to examine patterns of browsing and remote to understand how different
herbivore boards from cattle to like wild antelope interact with
these resources. And what we did with that project was that
we use a code that could tag the
individual pods on the ground. And then we tagged how many were

(10:24):
there at the initial time of the month versus how many were removed. And then
we ended up calculating a consumption rate from
that to see how fast and where the animals were
going towards. Since we did do it across four sites across
the entire river, we got to see, like, which parts of the
river were most likely consumed by which kinds of animals.

(10:46):
So that was super fun. So by integrating these field measurements
with quantitative analysis, we were able to identify which
environmental and structural factors most strongly influence the
plot production in these hybrid ecosystems. It was just a very
cool intersection of machine learning and environmental science. And of
like, we brought AI to the middle of the desert, which is super cool.

(11:09):
And after returning to Dartmouth, I was. Able to expand my
entrance in environmental monitoring through a separate
project as a research assistant. And in that role, the goal
was to track plant phenophases such as leafing,
flowering and fruiting, using time lapse and satellite
Imagery to study how the climate variability impacts

(11:31):
plant growth and production. And for this project specifically, I'm
working with a PhD student at Dartmouth who set up cameras again in
those four different sites and has pictures from every single day for
those four years, which is super cool, that we have a huge data set to
be looking at. But that has also come with some challenges.
It's always super unexpected. People steal cameras a lot

(11:53):
or don't like where they're placed. So then there has to be, like, you have
to keep track that those are constantly still there. And that's hard when
the college shooting is in at Dartmouth and then the pictures are in
Namibia. But we've worked with those challenges and these
images are going to be used with machine learning models to train the model for
image classification, phonology detection. And that work

(12:15):
has just really cemented my interest for combining AI, remote
sensing and ecology to generate insights that can inform both
science and resource management. Yeah, and I think that, you know,
everything that you've talked about. I know, because we talk about AI a
decent amount, you know, just both at Redox and on this podcast, but
it shows that there are so many innovations that you can do

(12:37):
with something like this. Like, people think of the basic things, like, I don't
know, create a schedule for me for my day or a
grocery list or budget list or other things like that. And, you know, on
the security side, we think of it in other, like, specific ways of
how we can, you know, maybe look through alerts a little easier or things like
that. But I think this is a good, I don't know, just

(13:00):
perspective to think about how we can think outside of the box with these
technologies. Like, there's kind of no limit at this point
with what things can do. So, you know, having your
perspective on some of the projects that we've been working on
has been really great because you come in with, you know, some other ideas and
other uses for these kinds of tools that we might not have thought about

(13:22):
before. Yeah, I think it's been super fun. And
interesting to look into AI, especially because I think a lot of
people's view on it is it can make my life easier. And yes, it can
probably schedule all your meetings or give you summaries of, like, really
long papers, but if we can scale these for better use, it
can make, like, the world a better place rather than just your daily personal

(13:43):
life. So I think it's very important, at least to me, to use
this AI to enhance the world in some way positively.
It has been so neat to hear about the AI that
you're involved in conservation has a big place in my
heart. So to hear AI being used, it's
just wonderful. And, and it gives me hope for the future of

(14:06):
conservation that there are people like you out there doing work like this,
because it really does matter. Of course, AI in healthcare
matters as well. Of course it does. But I'm very excited
to hear about the projects you've been working on in conservation. Join us
next episode as we discuss more security challenges
impacting healthcare and practical ways to address them.

(14:28):
Marina, do you have anything else that you'd like to say as we
finish up this episode? Thank you for having me on and.
Wanted to remind everyone to use the link in the show notes to share their
comments, ideas and feedback. And don't forget to lock the back
door.

(14:49):
Sam.