Latest / The Mental Health Evolution / Ep 37: Trusted Tools for Peer Workers with Naveen Raman

Transcript

Ep 37: Trusted Tools for Peer Workers with Naveen Raman · The Mental Health Evolution

0:00
(Transcribed by TurboScribe. Go Unlimited to remove this message.) This podcast is part of the Frederick Podcast
0:02
Network.
0:03
Learn more at ListenFrederick.com.
0:05
Welcome to Mental Health Evolution, a podcast about
0:08
what's changing in mental health and why it
0:11
matters.
0:12
I'm your host, Rachel Harrison, inviting you into
0:15
honest conversations with people from all perspectives in
0:18
the field, clinicians, tech founders, investors, insurance companies,
0:22
and all the folks in between.
0:24
Let's explore what's working, what's not, and what's
0:27
next.
0:32
Hello, everyone.
0:34
Welcome back to the Mental Health Evolution podcast,
0:38
where we are talking about how the landscape
0:41
is quickly evolving in the mental health industry.
0:45
Today, we are joined by Naveen Rahman, a
0:47
PhD student at Carnegie Mellon University.
0:51
Earlier this year, we talked about a growing
0:53
trend in behavioral health around using AI tools
0:57
to support triage and care navigation.
1:00
The idea is that these systems can help
1:02
people with lower acuity needs get connected to
1:06
information and resources more quickly, while also helping
1:10
to ensure that more complex needs are routed
1:13
to the appropriate level of care.
1:15
As we've been looking at what that actually
1:17
looks like in practice, we came across PeerCopilot,
1:21
a chatbot and AI-powered assistant being developed
1:24
at Carnegie Mellon University.
1:26
PeerCopilot is designed to support peer support workers
1:30
in behavioral health organizations.
1:33
One of its key design features is that
1:35
it does not draw from the open internet.
1:39
Instead, it uses a curated set of vetted
1:42
resources, meaning the information it provides comes from
1:46
pre-identified trusted sources rather than unfiltered online
1:50
content.
1:51
In practice, this allows peer workers to take
1:54
a conversation about someone's needs like housing instability,
1:59
benefits access, or connection to community supports, and
2:02
quickly turn it into structured next steps using
2:05
reliable information.
2:07
PeerCopilot is currently being tested in real behavioral
2:11
health settings, where peer workers are supporting people
2:14
through complex and often urgent needs with limited
2:17
time and resources.
2:18
Today, we are going to talk with Naveen
2:21
about how the system works, how it's being
2:23
used, and where tools like this fit into
2:26
the future of care delivery.
2:28
And as always, before we start talking with
2:31
our guest, we'd like to bring up some
2:34
relevant articles related to our guest topics today.
2:37
These may be helpful for listeners who want
2:39
to learn more about this topic and dive
2:42
deeper.
2:42
As usual, all articles will be linked in
2:45
the show notes for this episode.
2:47
So the first article I want to mention
2:49
is called AI Use in Mental Health Help
2:53
Seeking and Support.
2:54
And this research looks at how people are
2:58
already using artificial intelligence tools in mental health
3:02
contexts, including looking for emotional support, coping strategies,
3:07
and general guidance.
3:08
It highlights that while many people still prefer
3:11
human providers for care, AI tools are increasingly
3:15
becoming part of the mental health support landscape,
3:18
especially for individuals dealing with anxiety and depression.
3:22
The article also raises important questions about safety,
3:26
reliability, and the need for appropriate oversight, all
3:30
topics that we're going to dive into a
3:32
little bit more with our guest here in
3:34
a moment.
3:35
And then the next article is called Human
3:38
in the Loop, AI in Healthcare Systems.
3:41
And this article explains the concept of human
3:45
in the loop artificial intelligence, where AI systems
3:48
are designed to support human decision-making rather
3:51
than replace it.
3:53
It focuses on why this approach is especially
3:55
important in healthcare and mental health settings where
3:58
context, judgment, and lived experience are essential to
4:03
good care.
4:03
It also discusses how these systems can reduce
4:06
workload while maintaining safety and accountability.
4:10
And then lastly, article number three is about
4:14
the system we're talking about today.
4:16
It's titled Peer Copilot, AI Assistant for Behavioral
4:20
Health Navigation.
4:21
And this article introduces Peer Copilot, a chatbot
4:25
and AI-powered assistant developed at Carnegie Mellon
4:28
University to support peer-run behavioral health organizations.
4:32
It describes how the system helps peer workers
4:35
create wellness plans, navigate community resources, and structure
4:39
next steps using a curated database of trusted
4:42
information.
4:43
This system is designed to support and not
4:46
replace human decision-making and is being tested
4:50
in real behavioral health workflows.
4:52
I also just want to make a note
4:54
here that Peer Copilot is not associated with
4:58
any other tools that have the Copilot title
5:01
or brand to them.
5:03
It is a separate thing altogether.
5:05
So with that background, Naveen, we are really
5:09
excited to have you here to dig in
5:11
a little bit more.
5:12
Thanks for joining us.
5:13
You guys are having me here today.
5:15
Absolutely.
5:16
So I'd love it if you can kind
5:19
of explain in your own words what Peer
5:22
Copilot is, what it does for someone working
5:25
in a behavioral health setting.
5:27
Yeah, so Peer Copilot really came out of
5:30
this kind of collaboration we've been doing for
5:31
the past, say, couple of years with some
5:34
folks in New Jersey known as CSPNJ, this
5:37
kind of big statewide peer-run organization there.
5:41
And what we're looking for, or what they
5:42
were looking for, is a sort of tool
5:44
that can help them with their peer sessions
5:47
on this kind of day-to-day basis,
5:49
help find information on things like benefits, on
5:52
resources, and things like that.
5:54
And so what the tool does at a
5:55
high level is you take this kind of
5:57
existing language model, GPT, whatever you want, and
6:01
what you can do on top is you
6:03
can add certain what I'll call tools in
6:06
the kind of formal language on top.
6:07
What that means is you can add things
6:09
like a benefit navigator, where if I give
6:12
an information about a person's income and this
6:15
and that, it can tell you what benefits
6:16
they're eligible for.
6:18
Or if I give you a set of
6:19
documents, like, for example, from CSPNJ, I give
6:22
you a bunch of documents related to, say,
6:25
food or housing or whatnot, it can use
6:27
something known as sort of retrieval augmented generation
6:30
to search through these different documents and find
6:33
the ones that are most relevant.
6:34
The benefit to this is that it will
6:37
call these tools or use these kind of
6:39
additional features so that it can avoid relying
6:42
too much on its kind of background knowledge
6:44
that whenever it's kind of unsure or need
6:46
very specific information, it can rely on the
6:49
very specific information provided to the kind of
6:52
chatbot by us using these tools.
6:55
And so at a high level, it will
6:56
combine the best of kind of both worlds,
6:58
ideally, and it'll help peer providers during their
7:01
sessions navigate different situations, recall information, and just
7:05
help them hopefully get through some of the
7:08
more, say, mundane tasks of, hey, when is
7:11
this particular service open or what kind of
7:14
food banks are near me in Trenton, New
7:15
Jersey, so that they can spend more time
7:18
talking to folks there and just kind of
7:21
focus more on the other parts of their
7:23
job and less on the kind of information
7:25
retrieval and those parts.
7:27
Oh, OK.
7:28
I like that.
7:28
And it sounds like it's pretty location-specific
7:30
then.
7:32
Yeah, I think one of the things that
7:33
people really were emphasizing there is a lot
7:36
of the resources people need are very location
7:38
-specific.
7:38
So somebody in Newark would need food banks
7:41
that are in Newark and they wouldn't be
7:43
able to drive two hours to Trenton or
7:45
whatnot.
7:45
And so, for example, in the database we
7:48
have, which is given to us by CSB
7:50
&J, we kind of went through all the
7:52
kind of things or places that people had
7:55
in mind.
7:55
We went through, found the address and some
7:57
information on it so that if we search
8:00
a query like, hey, what are some food
8:01
banks near me in Newark, it'll go through
8:03
and try to find food banks that are
8:05
actually very close by, within a few mile
8:07
radius, and give the ones that are most
8:09
relevant or most proximate.
8:11
OK.
8:12
And is it just connecting to resources or
8:15
are there other tools available as well?
8:18
There's a bunch of other tools.
8:19
So I think we have a bunch of
8:20
tools kind of built in.
8:22
And for example, one of them is this
8:24
kind of benefit navigator.
8:26
And what that'll do is we kind of
8:28
hard code.
8:29
We say if you're eligible for, say, SSI,
8:32
then these are the kind of eligibility requirements.
8:35
And if you put in certain information about
8:37
a person, we can tell you whether or
8:39
not they're eligible based on these requirements.
8:41
And so if the model has some questions
8:44
about, hey, is this person eligible, it can
8:46
use those tools.
8:46
We also have some very basic kind of
8:49
tools or functionality built in, which are sort
8:51
of standard at this point with a lot
8:53
of language models.
8:54
Things like a calculator.
8:55
So you can do things like, hey, if
8:57
I earn $200 more next month, am I
9:00
still going to be eligible?
9:01
Models have gotten better at doing this now,
9:03
but it's still helpful to have these kind
9:05
of tools built in so it can be
9:07
guaranteed that it'll do well.
9:09
And then things like web search for things
9:10
like, hey, what is happening next week in
9:12
Newark?
9:13
Or what is, if you want very up
9:14
-to-date information, you don't want to rely
9:16
especially on the model's prior information or any
9:19
kind of stored database of information because they're
9:21
all static.
9:22
And so if I want to ask something
9:25
like, what is the address for this particular
9:29
place?
9:30
You can either search through the database and
9:31
try to find it there, or you can
9:33
just do a web search.
9:34
It can kind of automatically crawl the web
9:36
or search the web and find that information
9:38
that's most directly linked.
9:40
And I think we have probably a couple
9:42
of the tools built in, but these are
9:43
the kind of biggest ones that come to
9:44
mind.
9:46
Okay.
9:46
And how did this project get started?
9:49
I'm curious to hear a little bit of
9:50
the origin story.
9:52
Yeah.
9:52
So I'm a PhD student right now, so
9:54
I can tell you kind of the perspective
9:56
I'm coming from is sometime, I think in
9:58
July 2024, August 2024, my advisor was telling
10:02
me, hey, there's this organization up in New
10:05
Jersey that wants to work with us, something
10:07
AI related.
10:08
And so we did a little road trip
10:09
up there from Pittsburgh to a couple of
10:12
different offices in New Jersey.
10:14
I think in the end of July, early
10:16
August, sometime in that timeframe, and we were
10:18
just chatting with folks there to see what
10:20
is their experience with AI, what kind of
10:21
things could AI help with, what is your
10:23
kind of workflow?
10:24
And it was all very interesting information in
10:26
a very different domain than the one that
10:27
I was used to.
10:28
So it was cool to just see kind
10:29
of what are their needs, how can it
10:31
help them, and then from there it kind
10:32
of just blossomed and we just kind of
10:34
worked together for the past couple years from
10:36
there.
10:36
Okay.
10:37
Are there any specific learnings that have kind
10:40
of stuck out along the development process, like
10:43
anything that kind of went really wrong or
10:45
something that you learned and had to correct
10:47
for?
10:48
I think there's a lot of different things
10:49
that change over time.
10:51
I think one of the things that's very
10:52
interesting is, I think especially in this peer
10:55
kind of provider space, they're very big on
10:58
making sure that the tool is not a
11:00
replacement, but more kind of complimentary.
11:02
And so they want to make sure that
11:04
the information given to them by the tool
11:07
serves like a very specific need and helps
11:09
them kind of bring some different ideas.
11:10
They don't want it to kind of tell
11:12
them exactly what to do and things like
11:14
that.
11:14
So making sure the tool does that was
11:16
definitely kind of a back and forth process
11:17
of how do we design the tool?
11:19
How do we prompt it?
11:20
So on and so forth.
11:21
I think one of the other big things
11:22
is just the kind of speed of model
11:24
development.
11:25
And so for example, the paper you mentioned
11:27
before, the paper we wrote is online since
11:31
I think November and between November and now
11:34
we've re-ramped the tool a good amount.
11:37
And the reason for that is the way
11:39
you can develop models, say when we started
11:41
in 2024, is very different than the way
11:45
you can develop models now.
11:46
You can develop it in a much better
11:47
way.
11:48
The kind of baseline models are much better.
11:50
And so for example, this idea of like
11:52
tool calling was more new and not as
11:55
kind of well, let's say developed back in
11:57
2024.
11:58
And so we had to manually tell the
12:00
kind of agent or the kind of model,
12:03
hey, these are the things you should do.
12:04
And so if you look in the paper,
12:05
you'll actually see like a diagram that says
12:07
explicitly called benefits, explicitly called this and this.
12:11
But now the model can handle all this,
12:12
it can, it knows when to call which
12:14
tool.
12:14
And so we just have to give it
12:15
the appropriate tools and the appropriate rag and
12:17
whatnot, and it works a lot better as
12:19
a result.
12:20
And so I think along with that, it's
12:22
just people have more awareness of kind of
12:25
the power of AI, what can, what you
12:26
can do.
12:27
And so when we went in in 2024,
12:29
and I think most people there had never
12:30
really heard of like language models, had never
12:32
really used it.
12:33
But I think now it's definitely getting more
12:35
and more acceptance and people are seeing it
12:37
and just the potential for it is definitely
12:39
increasing just because the models themselves are getting
12:42
better.
12:42
Mm hmm.
12:43
They certainly are.
12:44
And I think you're right.
12:45
It's more and more widespread.
12:46
So can you walk me through like an
12:49
example of how this is being used, like
12:51
an example of a call where someone might
12:53
call into this group for peer support and
12:56
use Peer Copilot?
12:58
Yeah, so I think the way, so we're
13:00
still kind of trialing this out and seeing
13:02
how people use it.
13:03
But one of the many ways that people
13:05
talk about this is folks who really, people
13:08
who come in who really need some kind
13:10
of resource.
13:10
I think that's maybe the best kind of
13:13
use case we have in mind.
13:14
And what you can imagine is someone comes
13:16
in and they're saying, hey, I'm in Trenton
13:19
and I really need food at the end
13:21
of the day today, but I don't really
13:22
know where to go.
13:23
And I have some restrictions, like I really
13:26
want, I'd rather eat halal food because I'm
13:27
most of them or something like that.
13:28
And so they would, and then you could
13:30
type into the chat.
13:31
You could say, as you're kind of working
13:32
with them, you can say, hey, we're going
13:34
to look for food by searching Google's app
13:36
and then both them and the kind of
13:38
person they're working with, they both go on
13:40
the computer and they can see, can I
13:41
find food pantries that are good from halal
13:45
that are nearby or near Trenton?
13:47
And it'll list some pantries and then you
13:48
can kind of go through and click on
13:49
them and see what are the different results
13:50
and things like that.
13:52
But you can also do this in ways
13:53
that are not for resources.
13:55
So I think one of the examples that
13:57
we had during one of our trials that
13:58
we were testing out was something like you
14:00
can imagine a senior comes in and seniors
14:02
like, hey, I really want to stay active
14:04
and try out different things, but I can't
14:08
really like, I can't really run or do
14:10
anything because I got knee replacement.
14:12
And so, but what are the best exercises
14:14
I can do?
14:15
And so in this situation, you can kind
14:17
of brainstorm ideas with peer co-piloting and
14:20
ask, hey, this person cannot, they cannot run
14:22
too much or do too much high impact
14:24
sports.
14:25
What are some ways that they can stay
14:26
active and some good exercises, both for the
14:29
upper body and the lower body, and it'll
14:30
give you like a detailed exercise plan or
14:32
something like that.
14:33
And so they can kind of like have
14:34
a conversation saying, hey, these are some ideas.
14:37
What do you think about this?
14:38
And kind of go back and forth from
14:39
there.
14:39
I like that.
14:40
And so you mentioned, I think, in this
14:42
example that both the peer support person and
14:46
the person looking for assistance can both utilize
14:49
this tool and see the screen.
14:51
Is that accurate?
14:52
Yeah, I think it's very heterogeneous, depending on
14:54
how people want to use the tool.
14:56
And so some folks have said, hey, we
14:57
want to use this tool during the session
14:58
so that like the peer provider and the
15:00
service user both kind of talk and then
15:01
they can both see the screen and whatnot.
15:03
But there are other folks who have said,
15:05
hey, what I really want to use this
15:06
for is like research after hours.
15:07
Like, let's say that somebody brings up something
15:09
during conversation.
15:10
Maybe I want to do some more research
15:11
on this afterwards or before the conversation starts.
15:14
So the thing is really the heterogeneous use
15:16
case, depending on maybe where somebody is coming
15:18
from.
15:18
So I think it depends really on the
15:20
individual, depending on how they want to use
15:21
it there.
15:22
Okay.
15:23
And then I know that you have mentioned
15:24
that this is not the same as OpenAI,
15:27
but you also mentioned a feature about searching
15:30
the web.
15:31
So can you talk a little bit about
15:33
the safeguards in place and what was important
15:37
from the use perspective for that?
15:41
Yeah.
15:41
So I guess I'll talk about a few
15:43
things.
15:43
There's maybe two categories of ways you can
15:45
build language models in the same age.
15:48
One is you can build essentially what's called
15:50
like a wrapper around like an existing large
15:52
language model.
15:53
And so you can take something like GPT,
15:56
which is a lot, hundreds of billions, if
15:58
not even more at this point, number of
16:00
parameters.
16:00
And you can say, I'm going to make
16:02
what are called like API calls.
16:03
I'm going to send different requests or different
16:06
prompts to GPT, get some responses back and
16:08
do something with that.
16:09
The other way you can do things is
16:11
you can do what's called fine-tuning.
16:12
And what fine-tuning does is you can
16:14
take a smaller language model, but train it
16:16
on a big dataset that you have a
16:19
bit of access to.
16:20
And you can say, Hey, based on this,
16:22
how does the model change?
16:23
In this case, we're doing a TPT wrapper
16:24
rather than fine-tuning because we don't have
16:27
access to gigantic datasets here.
16:29
And the other kind of caveat is that
16:30
if you start fine-tuning models, it has
16:33
some kind of computational burden, which is not
16:35
terrible, but you have to kind of deal
16:37
with, you have to deal with smaller models.
16:40
And more importantly, you have to find a
16:41
way to actually serve those models, basically make
16:43
those models public in some way, which requires
16:46
infrastructure and cost.
16:47
When we build this GPT wrappers, the ways
16:49
we can do a lot of sort of
16:51
what I'll call like safety or whatnot is
16:54
you can make sure the prompt is good.
16:55
And so we do a lot of work
16:56
going back and forth with the users themselves
16:58
or the kind of peer providers to see
17:00
what should the prompt be?
17:01
What should it do?
17:01
And we leave, we give explicit instructions to
17:04
say, these are the things you should do.
17:06
To avoid things like hallucination, I think adding
17:07
in the web search tool is quite useful
17:09
to make sure that, Hey, if you're unsure,
17:11
call these tools.
17:12
And this is something we are explicitly in
17:14
code in.
17:15
And then the way this tool is deployed
17:16
right now is it's deployed as what's called
17:19
like an Azure server.
17:20
And so Microsoft has a service called Azure,
17:22
which runs on the cloud and can do
17:24
a bunch of things with hosting models and
17:26
running web servers.
17:27
And so there are also what I'll call,
17:29
I don't know the exact word for this,
17:31
but there are kind of safeguards built into
17:33
that as well, where you can avoid things
17:35
like overly harmful language.
17:36
You can set thresholds or filters on the
17:38
bat and whatnot.
17:39
And so these are also built into the
17:40
model or built into the kind of deployment
17:42
on the model so that we can avoid
17:43
any kind of bad behavior, any kind of
17:46
negative outcome that comes as a result of
17:48
this.
17:48
Of course, at the end of the day,
17:50
I think relying on models themselves for very
17:51
important tasks is a bit tricky just because
17:54
no matter how good your model is, there
17:55
are ways it will break.
17:56
There are ways you can potentially break it,
17:58
of course.
17:58
And so normally we will always have like
18:00
a peer provider working with this model so
18:02
that they are kind of an extra line
18:03
of defense.
18:04
They can say, Hey, this doesn't make sense
18:05
or this result doesn't make sense or whatnot.
18:07
So I think building and deploying these models
18:10
is really more about having these lines of
18:12
defense in rather than just saying, Oh, we
18:14
can build the best model possible and it'll
18:16
all work out from there.
18:18
I like that.
18:19
And to me, what I'm hearing you say
18:21
is, yes, there are lines of defense in
18:24
the design itself, but the biggest maybe line
18:27
of defense is that this is being utilized
18:30
with a human.
18:32
I think this is like an adage a
18:34
lot in like computer security where I think
18:36
what they call it there is like a
18:37
Swiss cheese model where you have all these
18:39
different lines of defense.
18:40
Every single line of defense has certain vulnerabilities
18:42
or certain holes.
18:44
The kind of goal that you want is
18:45
to make sure that all the holes don't
18:47
line up so that there's no path from
18:48
like one end to the other.
18:50
And so as a result, you can make
18:51
sure that whatever mistakes the model makes, the
18:53
human catches it.
18:54
The human catch can't catch something.
18:56
Maybe the model will catch it or the
18:57
additional safeguards we add on top or something
18:59
like that.
19:00
So we just want to make sure that
19:01
whatever we do, there are different ways to
19:03
catch mistakes.
19:04
And just another thing we're doing, actually, in
19:07
addition to that, I think the folks at
19:09
CSP&J are running some kind of AI
19:11
training session sometime in the next week or
19:14
two.
19:15
And so what they're doing there is they're
19:16
teaching people about how to prompt, what kind
19:18
of issues might come up when you prompt
19:19
and how to think about how to be
19:21
kind of cautious, but also optimistic when working
19:24
with these sort of models.
19:25
And so I think all of these things
19:26
really go together as different lines of defense
19:28
to make sure things work out well.
19:30
That makes a lot of sense because you're
19:32
not looking for the type of person to
19:34
just read off of a screen.
19:36
You're looking for the type of person who
19:38
has some kind of training working as a
19:41
peer support to be like, wait, that doesn't
19:44
sound right.
19:45
Exactly.
19:45
This is such a new tool that I
19:46
think having some degree of like AI literacy
19:49
or technological literacy is very important because it's
19:52
the way that model works or the way
19:55
it might hallucinate might look different than, say,
19:57
what a person might do when they make
19:59
mistakes.
20:00
And so just being aware of this, being
20:01
aware of these are the things that the
20:03
model can do.
20:03
These are the things that maybe the model
20:04
cannot do as well.
20:06
And try to tell the difference between the
20:08
two and just kind of being guarded whenever
20:11
you use these models because the model will
20:13
say things you might have to go back
20:14
and verify and this and that.
20:16
Yeah, that makes sense.
20:18
So this is a specific, like you have
20:21
a specific use case that you're using this
20:24
for.
20:24
Are there any ideas or dreams or plans
20:27
even to broaden the scope of where this
20:30
might be available for people to use?
20:33
I think that would be nice.
20:34
I think right now we're using this with
20:35
one specific organization, but I think in general,
20:38
I think there's two maybe lines of things
20:40
that at least I imagine personally.
20:42
One is expanding to other peer organizations.
20:45
I think we're very happy to work with
20:47
any peer organization who wants to use AI
20:49
in the workflow.
20:50
And this is kind of the tool that
20:52
hopefully can help them out with different kind
20:54
of things.
20:54
So we would love to kind of integrate
20:55
this there.
20:55
And then also just as a way to
20:57
maybe at maybe like a county level or
20:59
state level hub would benefit navigation and things
21:02
like that.
21:02
And so I know a lot of people
21:04
deal with like, hey, how do I apply
21:05
for SNAP or something like that?
21:07
And so maybe there's some version of this
21:08
tool or even just some version of some
21:10
language model out there where people could use
21:12
it for assisting with benefit navigation, this and
21:15
that, just because I think reading through the
21:17
benefit documents and understanding these things are very,
21:19
very complex.
21:20
But I think these language models, what they
21:22
do really well is if you give them
21:23
like a fixed source, like for example, you
21:25
give them like the fixed booklet on SNAP
21:28
eligibility, it's very good at parsing this and
21:30
breaking this under simple terms and understanding this,
21:32
which can hopefully kind of lower informational barriers
21:35
and make it easier for people in these
21:36
kind of spaces.
21:37
To see what's going on.
21:38
Yeah, yeah, I can see that.
21:40
Interesting.
21:40
It occurs to me like a lot of
21:42
locations have a 2-1-1 resource.
21:45
Are you familiar with that?
21:46
Yeah, I think I've heard about 2-1
21:48
-1 before.
21:49
Yeah.
21:49
So it's similar.
21:50
It's usually like a hotline sort of thing
21:53
where people might be calling with because they're
21:56
in emotional distress, people might be calling looking
21:58
for resources.
21:59
But the need for resources is inherent in
22:02
that kind of support system and thinking this
22:07
could be a great utilization potentially for those
22:10
types of organizations.
22:12
Yeah, yeah.
22:12
I know whenever we have some queries, I
22:15
know 2-1-1 is like a resource
22:16
we have even embedded into the kind of
22:18
database there.
22:20
OK, so a lot of the times if
22:21
we have a query and it's something like,
22:23
hey, where do I find this food resource?
22:25
It'll tell you some of the food resources,
22:26
but it'll also say maybe you should contact
22:29
2-1-1 to find additional information or
22:31
something like that.
22:32
So maybe that's definitely another use case for
22:34
it.
22:34
I think there's a lot of different use
22:35
cases where we need to find certain amounts
22:37
of information or whatnot, and people could use
22:39
it there.
22:40
Do you ever think that it could be
22:42
used without that human interaction?
22:46
You're saying directly by, say, like a service
22:48
user?
22:49
Yeah.
22:50
That's an interesting question.
22:51
I think there are definitely risks, but also
22:52
benefits.
22:53
The benefits is if I need something urgently,
22:55
then I can kind of quickly go to
22:56
it and find something.
22:58
The risk is that maybe the tool is
23:00
not fully designed for that just because there
23:03
are, of course, things that might hallucinate or
23:05
say the wrong thing here and there.
23:06
I think it's also we're trying to build
23:08
in a few other things into the tool
23:10
to make sure it's more accessible so that
23:11
if we want to kind of go directly
23:13
to service users, then it could.
23:16
But I think it's a bit tricky just
23:18
because we don't want to ruin or even
23:20
tamper with the peer provider or service user
23:22
relationship.
23:23
And so this is why we're sticking mostly
23:24
to kind of helping peer providers for the
23:26
time being.
23:27
And then if people are interested, we could
23:29
potentially go directly to service users.
23:31
But I think it's a lot of complexities
23:33
that end up coming there.
23:34
For example, things like if my service user
23:37
speaks entirely Spanish, then does my tool do
23:40
well in Spanish?
23:40
If my service user is kind of either
23:42
hard of seeing or hard of cannot read
23:44
or something like that, can my tool kind
23:47
of say things out loud and whatnot?
23:49
These are things that we've tried to build
23:51
in.
23:51
Like, for example, it should natively support other
23:53
languages just because that's how language models work.
23:56
And I think we're planning to build in
23:58
like text-to-speech or things like that.
23:59
But I think there are all these kind
24:01
of more practical complications that start coming in
24:04
if you start thinking about, oh, we're going
24:06
to go directly to service users.
24:09
I agree.
24:10
Yeah, I could also see that when you
24:12
talk about this systems navigation, it does make
24:15
me wonder if there's a use case for
24:17
like Department of Social Services, right, to be
24:20
able to easily provide a way people can
24:24
quickly see, like, do I qualify?
24:26
Should I even make the appointment?
24:27
Like, I can see so many kind of
24:30
possibilities there of like, yeah, you qualify for
24:33
SNAP.
24:33
You should make this appointment versus like, nope,
24:36
your income doesn't match.
24:38
Don't waste your time.
24:40
Exactly.
24:41
I think that could be a very good
24:42
use case for this where I think it's
24:44
very good at like if you give it
24:46
fixed information, which ideally things like SNAP are
24:49
at least hopefully somewhat fixed where you can
24:51
tell this is what happened, then it can
24:53
just process your information in whatever format you
24:56
give it and tell you what you qualify
24:58
for, what information you need.
25:00
It's kind of a more flexible version of
25:02
the calculators that are already out there.
25:05
And so I think that this kind of
25:06
thing could definitely be useful and wouldn't be
25:09
too complicated to build this because it would
25:11
just need to be like a minorly prompted
25:13
or tailored version of GPT or kind of
25:16
a minorly modified version of whatever we have
25:18
right now with PureCopilot.
25:20
So that's the hope there.
25:21
OK, I love that.
25:23
Is there a way for organizations to get
25:26
connected with you if they are interested in
25:29
being a test case or utilizing this software?
25:34
Yeah, of course.
25:34
I'm very happy to chat with any organization.
25:37
If you look at the paper itself, it
25:38
should have my email, I believe.
25:40
And so or you could just search my
25:42
name, Naveen Rahman.
25:43
I'm sure you can find my email somewhere
25:45
on the Internet.
25:46
And I'm happy to if you just send
25:47
me an email, I'm happy to chat with
25:48
any organization who has questions about the tool,
25:51
even if they have questions about AI, I'm
25:53
happy to chat with them and see what
25:55
can AI do for them.
25:57
Even if it's not necessarily PureCopilot, even if
25:59
they want to use another tool, I'm happy
26:01
to see what can AI do in the
26:03
space for them.
26:04
I love it.
26:05
I love what you're doing.
26:06
I really appreciate the intentionality, the focus on
26:10
supporting the organization that you're working with, the
26:13
human line of defense.
26:14
All of those things are great.
26:16
If you could leave our listeners with one
26:19
last thought or maybe a most important thing
26:23
that you think they know about AI models
26:25
or integrating this in the mental health space,
26:27
what would that be?
26:28
Yeah, I think maybe the thing I would
26:30
say is AI models are very powerful, but
26:32
they also need to be very kind of
26:35
specific or tailored towards a domain to make
26:38
sure that they work in this specific context.
26:41
And so just thinking about both what can
26:44
AI do that might be able to help
26:46
me, but also how do I make sure
26:47
that it works within my context so that
26:49
it works with the people and it works
26:52
within those relationships in the kind of right
26:54
way that I want it to work is
26:56
both a big question, but also potentially very
26:58
impactful.
26:59
Yeah, yeah, awesome.
27:02
Well, Naveen, I really appreciate you taking the
27:04
time.
27:05
You and your team are doing some great
27:07
things.
27:08
So thank you for that.
27:09
We will definitely have the article, the paper
27:12
that you all wrote in the show notes.
27:14
We'll also put your email.
27:15
It sounds like you're OK with that in
27:17
the show notes so people can reach out
27:18
to you.
27:19
It's been a great, great conversation.
27:21
So thanks for being here.
27:23
And of course, thanks for having me and
27:24
thanks for the chat.
27:25
Ah, so we will be back next week
27:28
on this pod to discuss more about issues
27:31
relevant in the mental health care community.
27:33
Thank you for listening and bye for now.
27:37
Thanks for listening to the Mental Health Evolution
27:40
podcast.
27:42
Join the conversation.
27:43
Find us on socials, send us your thoughts
27:45
and tell us what you're seeing in your
27:47
corner of the mental health world.
27:49
Until next time, stay curious and keep evolving.