Latest / Agentic Conversations (formally mlops.community) / AI in Healthcare // Eric Landry // #249

Transcript

AI in Healthcare // Eric Landry // #249 · Agentic Conversations (formally mlops.community)

0:00
OK name is Eric Landry, CTOCAIO is a Te'o health and I love my
0:08
coffee. Americano, my espresso machine,
0:13
is one of my prized possession in my house.
0:16
Welcome back ML OPS community, we are here with another
0:19
podcast. I am your host as usual, Dimitri
0:22
OPS and we talked all about the healthcare space, bringing AI
0:27
into the healthcare space. Hi, high high value uses of a
0:34
rag chatbot. Not this HR bullshit that you
0:37
get high value use cases and the nuances of it, the ethics of it,
0:45
the conversational design and how it can be so difficult.
0:49
I really appreciated he's been doing AI for so long in
0:55
healthcare and I love the fact that he still is hands on
0:59
building. You can tell when we talked to
1:02
him, he's going in there, he's hacking around.
1:05
It probably would have been easy for him to sit back and become a
1:09
manager, become an executive by now.
1:11
But no, he's having fun with it. So let's get into this
1:15
conversation and as always, if you liked it, make sure to tell
1:20
one friend so we can keep this gravy train rolling.
1:32
So I think we should start with what we were just talking about
1:39
and that is you coming into this field in 2005 and explain to me,
1:45
paint the picture of what things were like trying to do ML and AI
1:51
in 2005. Yeah.
1:54
So I entered the field through my school work at University of
2:00
Texas. And back then my studies were,
2:07
my thesis was about document clustering of scientific
2:13
articles about genetics. My experimentation was basically
2:18
testing different algorithms. A lot of it is, you know,
2:22
processing the data, you know, an LP is usually pre compute
2:29
intensive. So finding the compute and the
2:33
resources and the memory was always kind of constraint.
2:39
I was at that time working at Sun Microsystems.
2:42
So you know, I at home I had a little MacBook and I would
2:48
start, I would start the algorithm running overnight and
2:53
then half the time wake up in the morning with a stack
2:55
overflow or some other crash in my computer.
3:01
So eventually I want to take in a lot of my work, battle a lot
3:05
of the school work with me to work at some microsystem that we
3:09
had a lab full of the latest greatest servers.
3:12
So, you know, what would take overnight at home would take a
3:15
couple hours. Over lunchtime, it still took a
3:18
couple hours, right? But, you know, and then all the
3:23
the algorithms, as I'm had mentioned before, I wrote in
3:28
Java. So there was a lot of attention
3:31
paid to you understanding, you know, reading the research paper
3:35
and then, you know, writing the proof in Java.
3:39
And a lot of times I would find it through, you know, that the
3:42
proofs were actually incorrect in the papers.
3:45
Not a lot of times, but sometimes.
3:48
And so, you know, being able to, you know, Fast forward, you
3:53
know, being able to like iterate through different algorithms
3:58
quite easily was kind of a mind blowing to me.
4:04
But yeah, I just kind of get strains around the compute the
4:07
memory just, you know, data processing, data transformation
4:14
and cleaning, etcetera was all a lot more challenging back then.
4:18
You have. People don't realise how good
4:21
they have it with Pytorch these days.
4:23
Yeah, I got real good at, you know, in Java figuring out what
4:27
the the optimal data structures were and how to how to optimise,
4:33
you know, data compression and and things like this to achieve,
4:37
you know, some some of the algorithms you need everything,
4:39
you know, all the data and memory.
4:40
So that was a lot of tonnes of challenge.
4:43
So, yeah, it feels like you throughout your career have been
4:47
attracted to very hairy problems because now you're in the
4:53
healthcare space working in AI, and there's a lot of constraints
4:58
and there's a lot of difficulties in healthcare,
5:00
whether it is with the data itself or it is with just how
5:05
you can use AI, the different use cases.
5:09
And so I feel like we're going to touch on a lot of these
5:13
different pieces, but it would be nice to know about what you
5:18
were doing at your last job, Babylon Health, because I know
5:24
that you were the director of AI Engineering and the Conversation
5:28
Platform that was like the prequel to what you're doing now
5:32
at the current job. And what did that look like?
5:35
Those are two different roles. The AI platform, my team was
5:40
building a a cloud agnostic platform because at that time we
5:47
were attempting to be cloud agnostic.
5:49
So we built the infrastructure around, you know, being able to
5:54
deploy our models regardless of, of, you know, what it is Azure
5:58
or AWS or what not. So all the challenges along, you
6:04
know, how to optimise the, the, the different types of
6:08
deployments. We had streaming, inferencing on
6:12
a stream, Kafka stream. We had inferencing as a service
6:17
and then, you know, batch type streaming.
6:19
So we were supporting all those. We're also supporting, you know,
6:23
cloud based training. And these were NLP use cases
6:27
still. There was any, it was all that
6:29
like it was NLP, it was, you know, structure data.
6:33
We had a lot of of medical data we were training models on.
6:42
So that was that was that piece. And then also the conversation
6:47
AI, you know, conversational platform, we built chat bots.
6:52
They're intent based chat bots. You know, back in the day, I
6:55
just need anybody to remember intent based chat bots.
6:59
So we were using, you know, the source frameworks to build the
7:03
chat bots. Unfortunately, this we were six,
7:08
we were building a conversation platform which included live
7:13
chat as well as, you know, automated responses with the
7:17
chat bot and you could hand off that, you know, the user could
7:20
request to be handed off to a live agent if they weren't
7:24
satisfied with response from the chatbot.
7:27
So we built the whole system around all those features and we
7:32
unfortunately we released our chatbot probably three or four
7:36
months before Babylon closed shop in the US.
7:41
But it, our early indications that we were, our whole, whole
7:46
focus was to deflect users patient or our users from the,
7:55
the live agents with an automated agent.
7:59
And we were able to deflect up to 40% of the users.
8:04
So, you know, that's the real use case and the real value of
8:08
of those chat bots. Yeah.
8:10
And again, this is something that it's like history rhymes
8:14
right now. Everybody understands that
8:17
because I think everybody's tried to throw a support chatbot
8:21
with an LLM at their her product.
8:25
Yeah, it's interesting to kind of see the progress so far and
8:31
kind of trying to I guess get through the hype and really
8:37
understand what the real value is.
8:40
I'm I'm kind of slowly evolving. My thinking is started to evolve
8:45
where you you can't just release an LLM based chat bot without
8:50
appropriate controls. And so the reality is almost
8:54
something kind of in between maybe I don't know, I'm not
9:00
quite there yet, but I'm experimenting with some
9:04
technologies that would enable kind of AI guess a hybrid type
9:10
chat assistant and. When you say hybrid, you mean
9:15
like rules based or intent based plus LLM?
9:17
Yes, that makes sense. Yeah.
9:20
You can't let it be anything or else you wind up on the front
9:24
page of the Internet. Yeah, because yeah, exactly.
9:28
Yeah, you said something that chat bot said something it
9:31
shouldn't have or it is just basically a wrapper for GBD
9:36
Ford, Yeah. I mean, and understanding
9:39
healthcare, you know, OK, so somebody sold a truck for a
9:42
dollar like healthcare, the stakes are a lot higher, right?
9:47
So we kind of have to get it right.
9:49
Yeah, and that goes back to the hairy problems that I was
9:53
talking about. Like there's real ethical
9:56
issues. There's human lives potentially
9:59
at risk when you're dealing with your use cases as opposed to.
10:03
I do understand that selling a truck for a dollar is a bummer,
10:07
but it's not like there's human lives at risk, right?
10:10
Yeah. Yeah, exactly.
10:11
I have. That's the the punchline I used
10:14
to tell my teams that when I was in travel at Expedia, it was,
10:20
you know, engineers want to tend to engineer and a lot of times
10:26
over engineer things. And so my punchline a lot of
10:28
times was, well, you know, if this thing breaks, planes aren't
10:33
going to fall out of the sky. You know, I had to rethink that.
10:37
In healthcare, you know, the stakes are a little bit higher,
10:41
you know, than running somebody's vacation.
10:43
Exactly. At Expedia, it's like, well,
10:46
somebody's vacation, maybe you're not even ruining it.
10:49
Maybe you're just giving them an opportunity to do something else
10:52
more interesting. But yeah, with healthcare, it's
10:54
a different story. And going back to some of these
10:59
difficult issues with data, I mean, we all probably are
11:05
familiar with PII and how big of a pain in the butt that is.
11:10
So that's like, First off. But I remember when I was
11:13
talking to a friend that I mentioned before who was working
11:16
at Babylon Data, Jeremy, and he was saying when I talked to him
11:21
back in 2020, he mentioned how one of the engineering issues
11:26
that he was looking at in that moment was how to keep all of
11:32
the data that is created in Canada in Canada.
11:37
They don't. He didn't want anybody else to
11:40
be able to touch that because that's illegal.
11:42
But he still wanted all of the ML engineers to be able to train
11:47
their models and have the most robust data sets that they
11:53
possibly could have without like jeopardising that Canada data
11:58
set. And so those are some, yeah,
12:01
like data access issues that are not easy to figure out.
12:05
Yeah, that's a real challenge. And you know, the issues that
12:09
you cannot transfer data between regions.
12:11
So you can't transfer data from the UK to the US to train a
12:15
modeller or a, you know, inference, you know, on a model.
12:19
So you have to come up with strategies around how to keep
12:23
the data in place while you developed your solution.
12:28
You know, a lot of times we had different, you know, sometimes
12:31
we had different models, you know, and, and a lot of times it
12:34
kind of makes sense because the data is different and it's a
12:38
different shape and different, different rules that can be
12:42
applied to it. We also had experimented with
12:47
Federated learning where the data stays in place, but you
12:52
train the model and so you're not transferring the data
12:55
between regions, but you're, you have agents, I guess you would
12:59
call it agents that live in the different regions.
13:02
And then the the what's transferred are the weights of
13:07
the the model to, you know, come up with the optimal train model.
13:13
And did that go forward? I I think you mentioned that
13:17
wasn't the most successful POC it.
13:20
Didn't go forward. I think they're, you know, in
13:24
these enterprises, different things take priority that that
13:29
actually never surfaces a priority.
13:31
But it was, it was a great experiment and you know, we were
13:33
quite pleased with with the work we've done, which is pretty cool
13:37
because it also, it's not just between regions, but you can,
13:41
you know, train the model on your phone.
13:43
You know, since you got, you know, 100 users, you know,
13:47
thousands of users, they can, their data never leaves their,
13:53
their device. But we can train a model right
13:56
based on all that data. Which is much more appealing for
14:00
things like healthcare or if it's like things that I'm saying
14:05
with my therapist, I would prefer that data to be my data
14:09
and not have to send it out to get models trained and whatnot.
14:14
Yeah, our POC was on skin cancer images, you know, images that
14:20
were taken. We tested it on on a mobile
14:24
device training, you know, to detect if there's skin cancer or
14:27
not. So you can see, you can imagine
14:30
the uses for this kind of technology.
14:33
Yeah, 100% and. You.
14:36
Did bring up something there on how it didn't really become a
14:39
priority. In your experience leading AI
14:44
and ML teams, what are the biggest priorities and how do
14:49
you champion for them? How do you know which ones
14:54
should be going to the top of the line?
14:57
And like being able to, with confidence, say, yes, here's the
15:01
metrics that we're looking at. We know that if we're doing
15:04
this, we can have success. Yeah.
15:08
You know, a lot of these priorities come from the top
15:11
down, right? So you got, you know, product,
15:14
product people, you know, basically doing the the market
15:19
research and understanding where the best use of the resources
15:24
are. I always treated it as we were
15:28
equal partners in it. And that, you know, I've been in
15:33
the industry for a long time. I'm not a product manager, but I
15:37
built a lot of products, right? So, you know, together, you
15:42
know, if you have the right relationship with a product
15:44
manager together you can decide what priorities are.
15:49
And you know, the challenge is how to balance that with you
15:52
want your team to feel you know and able to experiment and you
15:59
know, explorer creative solutions.
16:03
So how do you balance what the enterprise want you to do versus
16:08
what you think you can do and what you can build?
16:12
So it's a tricky balance. And you know, of course, you got
16:14
to play some politics to get through it all.
16:16
And at the end of the day, hopefully it all works out.
16:19
Your team, you know, I always felt like my team should have
16:22
agency to decide what they build, you know, within
16:27
boundaries, right? But it can work in a reverse
16:32
too. You know the engineering team
16:33
should be able to propose the new products and features,
16:37
right? Yeah.
16:39
So what are you working on now? Currently working in a startup
16:43
called Zateo Health. The, our focus is on bringing
16:49
healthcare to underserved communities and what that the,
16:54
the, the kind of the why behind that is that they're so, so we
17:00
want to give the, these communities information about
17:05
healthcare to keep them engaged in their healthcare.
17:09
The studies, their studies out that show people who are engaged
17:15
in their health tend to live healthier lives, right?
17:19
And the reason why it's important in these communities
17:22
is because these communities like black and Hispanic
17:25
communities, at least in the US, they don't trust a healthcare
17:28
system. I think something like 50% of
17:32
blacks and Hispanics don't trust the healthcare system.
17:35
And this kind of lack of trust in the system leads to them not
17:39
being engaged in their healthcare.
17:42
There's a study I quote often it's posted on National
17:48
Institute of Health that says there, I think it was a year was
17:52
2018, there were 100,000 preventable deaths at the cost
17:56
of $100 billion to the healthcare system because people
18:01
were not engaged in your health, not they're basically weren't
18:05
taking their medications. So to me this feels like pretty
18:09
low hanging fruit. You know, everybody wants looks
18:13
for shiny objects. You know, how do we solve
18:15
cancer? What not like, while that's
18:17
super important, but this looks like to me, keeping people
18:22
engaged in your health seems like a solvable problem, right?
18:27
So that's what our kind of focus is on.
18:31
And the technology we're building is basically what I'm
18:36
calling conversational AI framer, kind of the lead
18:40
technology, of course, is a rag based chatbot, but it's built
18:47
within a system. So, so I'm thinking about it
18:50
holistically about what can these conversations tell us not
18:55
just about individuals, but what could they tell us about certain
19:00
populations and segments within the populations.
19:03
And so we can, from this, we can kind of begin to make inferences
19:09
about how we can keep these people healthy, right?
19:13
Not just kind of in place like statically analytics, which we
19:20
will do on these conversations, but but in real time, what, what
19:25
are the trending topics? What are the trending terms and
19:29
what can we learn from that and how can we action action it?
19:33
So just a simple example in Austin, TX currently, you know,
19:40
so we can build a window of we said I used to do this at
19:44
Expedia. We had, we were using elastic
19:47
surge to make real time inferences about certain things
19:55
and the way you do that is put one window against another in
19:58
time. So the last week compared to the
20:02
the last month, what are the trending topics or trending
20:06
terms that stand out? So you can imagine in
20:09
healthcare, you know, in Austin, TX this week, coughs and fevers
20:15
are trending and it happens to be flu season.
20:20
Maybe we should send a notification to the people in
20:23
Austin that they should go get their flu shot, right?
20:27
That's a very simple example, but you can imagine the kind of
20:31
features and things that we can we can create out of this.
20:36
So I understand correctly engaging in your health is
20:41
primarily just going to the doctor.
20:45
It's taking medication that's prescribed to you.
20:48
It's and it's that kind of stuff that some of us might think
20:52
like, yeah, of course that's what you do.
20:54
But you're saying that that's not the common thing that some?
21:00
People do, you would think, but yeah, I mean, you know, some
21:06
people have, you know, for example, diabetes, you know,
21:10
there's a daily, you know, you need some people need a daily
21:13
reminder to, you know, test their blood glucose levels.
21:17
You know, some people need a reminder, you know, even me,
21:21
like I I generally stick to my annual check UPS, but you know,
21:28
it slips like I usually do it in the summer, but guess what last
21:32
year I did in December. You know, who knows what could
21:36
happen in those few months, you know, but it's this just kind of
21:41
engagement, not just in kind of like daily thinking about your
21:45
health, but also kind of annual checkups if you have for, you
21:51
know, prostate cancer. There's a the notion of people
21:55
who have been diagnosed with prostate cancer, there's a
21:59
notion that some of them can be, you know, the treat.
22:03
There's no treatment. They're just monitoring their
22:05
health and they need to stick to the plan, right, to keep them
22:09
healthy. OK.
22:12
And. I have been known to talk a
22:16
little bit, a bit of shit on rag chat bots.
22:20
Yeah, because I think they, for the most part, aren't fully
22:27
thought through and the use cases are not that high value in
22:33
your use case. If it is keeping people engaged
22:37
and getting them to take their pills or go to the doctor and
22:42
it's saving lives, like the stakes are very high there.
22:45
I think that is a great use of a chat bot.
22:49
So you just scratch an itch right there.
22:52
Like, yeah, I'm with you, man. Like, there's so much hype
22:57
around this. There is value that's
23:01
undoubtable, but there's so much hype.
23:05
And I feel like there's a whole lot of attention paid to, you
23:10
know, different types of retrieval, all, you know,
23:16
different, you know, lane chain and all these technologies that
23:20
I feel like we haven't really figured out how to get the most
23:24
out of what we have. Let's step back and understand
23:29
what it is we have and how we can apply it to a real use case.
23:36
I'm with you that like people think it's magic, man.
23:40
It's not like, and, and we discussed earlier about
23:44
guardrails. There have to be guardrails,
23:47
especially in something as consequential as your, your
23:50
health, right. You know, we're primarily, you
23:54
know, seeking engagement with our population through, you
23:59
know, notifications and information.
24:02
You know, a lot of my experimentation has been around
24:05
how to limit hallucinations, how to treat people and engage them
24:14
in a conversation where they are curious about their health.
24:18
You know, how do we lead them in a conversation about their
24:21
health? We have, we have healthcare
24:23
providers and professionals, owners aren't on our staff.
24:28
Then I talked to them and it's, it's really interesting to me
24:31
how doctors think about when somebody walks in the door, how
24:35
do they size them up and how do they understand how best to help
24:39
them and support them in their health.
24:42
And so a lot of the experimentation I've been doing,
24:46
aside from how to limit hallucinations is how do you
24:50
engage them in a conversation, especially these underserved
24:54
populations that don't trust the healthcare system, how do we
24:58
get, you know, bring trust to them?
24:59
So first of all, our, our knowledge base is a curated
25:03
trusted knowledge base. So any information we give them
25:07
is from that knowledge base and limiting the knowledge.
25:12
So basically, I'm not using the LLM for its knowledge.
25:15
I don't want its knowledge. I may as well go and search
25:18
Google, right? I want the knowledge for a
25:22
solution to come from the knowledge base of our curated
25:26
information, the high quality data, high quality data, right
25:31
that our healthcare professionals have advise us on,
25:34
right? And how do we give them that
25:38
information in a way that's meaningful to them?
25:42
It doesn't like, you know, their eyes don't gloss over when we
25:46
give them a whole bunch of technical information are are
25:51
the doctors on our staff and I've talked to they basically
25:55
say give them information that they need a bit at a time,
26:01
right. So you don't, you know, somebody
26:02
asked like, how do I get tested for prostate cancer or how do I
26:06
get tested for diabetes? You don't give them three
26:09
paragraphs. You give them the basic
26:12
information. There are three types of tests
26:14
you can do for prostate cancer. Allow them to then ask more
26:18
detailed questions about, you know, the PCA tests and then
26:23
leave them through it, you know, kind of prompt them to the next
26:27
question to ask that's more meaningful to the the patient
26:32
than just giving them three paragraphs, you know.
26:36
Yeah, try and make it bite sized and easily digestible, yeah.
26:41
And that's to me, that's that's a lot of the challenge.
26:44
You know, a lot of what I see, the information I see is like,
26:49
how do you respond to a question?
26:50
Well, cool. Like we have a really good way
26:54
to optimise question and answer, but how do you optimise the
26:58
conversation, right? So think of it as a
27:00
conversation, not a just a question and an answer, right?
27:05
Yeah, and that's where the theories on how much to give
27:09
someone come in. And you don't want to just give
27:13
them this blast of information, which is hard with LLMS because
27:17
they're constantly being super verbose.
27:20
So, yeah, so that's that's the challenge.
27:22
That's actually also sometimes I've done a lot of
27:26
experimentation on is tone of voice, like how you speak to
27:29
them. I was going to ask because of
27:32
the different colloquialisms in the underserved communities or
27:40
underrepresented communities. Are you trying to put on
27:45
different slaying or accents or is that kind?
27:48
Is that like a step too far? Is it seen as cheesy?
27:53
Have you experimented with it at all?
27:56
Yeah, not not. In that sense, what I've
28:00
experimented with is how to detect basically who that person
28:05
is, right? Not necessarily what we should
28:09
have a we intend to have a profile of the user like, so
28:13
we'll know their race, we'll know their age.
28:16
But how you know, you can make inferences about not just level
28:23
of education, but how, how they want to be conversed with,
28:27
right? And we don't right now, our tone
28:31
of voice is basically a standard tone of voice represented by our
28:37
branding, right? The the ZTEO branding.
28:41
But we intend to do experimentation about, you know,
28:45
how we converse with people. Like do you you do you want to
28:49
talk with yourself? Like, you know, do you want to
28:52
talk to somebody that has the same tone of voice as you?
28:55
Is that how you best respond and how you are best engaged in a
28:59
conversation or somebody opposite, you know, I you know,
29:03
that's kind of experimentation that that we want to do, right?
29:07
Yeah, I wouldn't trust it if it was talking like me.
29:11
It would be like probably. Most people would probably be
29:15
freaked out by talking to themselves, you know?
29:18
Yeah, completely. But that's fascinating to think
29:21
about like it. Does it give you more
29:24
credibility depending on the tone?
29:27
And if so, do you want to be more formal because that's going
29:30
to give you more credibility, or do you want to be more informal
29:34
and speak like people do from the different accents or
29:41
communities that they're coming from?
29:44
Yeah. Exactly.
29:45
And how? How do they best respond?
29:46
You know, maybe, you know, somebody with, you know, without
29:51
an accent, let's say a Hispanic accent responds better to
29:54
somebody with an American accent or something else, you know, who
29:58
knows, right? That's that's kind of the
30:00
experimentation that we we need to do, right.
30:04
And how are you evaluating this? Yeah.
30:07
So I've built a framework for evaluation based on Ragas I'm
30:15
using, I've done an integration with MO Flow to track the
30:19
experimentation. So baseline we, you know,
30:22
certain things that we're looking for, you know, obviously
30:25
a lot of the, you know, faithfulness for detecting
30:30
hallucinations and bias. What I found is that, you know,
30:34
a lot of these kind of basic bias detection mechanism don't
30:40
test, don't don't inform all bias.
30:44
You know, like the I found a data set on hugging face.
30:49
It was basically labelled set for bias against different
30:53
cultures, right? So I tested some of those
30:57
utterances against an LLM and and then the bias detection and
31:01
it didn't, it didn't catch some fairly heinous things that were
31:05
spoken. You know, that's so.
31:08
So I wind up writing some, you know, my own bias detections for
31:13
those cultures, right? And that's just hard coded in
31:16
there as guardrails. Yeah, it's not, it will be.
31:21
But we what I do, my process is that I'm testing a guest, a
31:28
ground truth set of data I have and before I deploy I do a test
31:34
to make sure it's good against that set.
31:37
I will also be monitoring in production so we can send alerts
31:42
if somebody starts screaming crazy shit like.
31:46
You know, so funny you say that because I was at a conference
31:52
last week, or AI quality conference, and somebody showed
31:56
me a screenshot of a support bot interaction that their company
32:03
was having. And the person literally was
32:07
writing in all capital letters, the fucking human, you dumb fuck
32:12
computer. And then the LLM answered like
32:17
something that was not getting them the human.
32:20
And that just infuriated the customer, right?
32:24
It's like the last thing you want to do is not get them the
32:28
human and they're just going ape shit on it.
32:32
And so yeah, that's what you want to try and avoid.
32:36
And I like the fact that you set up monitoring, so in case that.
32:42
Does. Happen and something slips
32:44
through the cracks, you know, right away.
32:45
And you can presumably you, you can step in and say, hey, sorry
32:49
about that. Let's take a few steps back.
32:53
Yeah, yeah. I think they're also, I'll add
32:56
that I don't know if I've mentioned it yet, but the, you
33:00
know, there I kind of view it as at least two types of bias.
33:03
There's this kind of like subjective type of bias, which
33:07
is, I know what we're talking about.
33:09
I mean, some of it's obviously it's not that subjective, but
33:12
you know, cultural sensitivity kind of things, right?
33:16
But there's also bias, like objective, clearly objective
33:20
bias. For example, you know, the
33:23
medical community. Like.
33:28
Advises the general population to get start getting tested for
33:32
prostate cancer at the age of 50.
33:34
I keep saying prostate cancer because our our MVP is going to
33:38
be basically focus on prostate cancer and we'll build out from
33:43
there for different types of disease.
33:46
We don't want to inform a black male to start getting tested at
33:52
50 because the guidance is is 40, the age of 40 to 45 because
33:58
they have a much higher incident.
33:59
The black and Hispanic communities have a much higher
34:01
incidence of prostate cancer. So, you know, there's this kind
34:05
of objective measurements that we have to take also, you know,
34:09
you know, data quality type measurements, it seems like you.
34:15
Decided to play the game on hard mode.
34:17
Not only is healthcare, the healthcare space, very
34:21
complicated and very hard, but now you're working in a space
34:27
that is tailoring to these underserved communities.
34:32
And there's a lot of difficult areas in that too that you have
34:35
to navigate through. Yeah.
34:38
I think, yeah. So a little bit about that for
34:41
me. You know, when Babylon shut down
34:46
last summer, I was kind of in a burnout, you know, the the the
34:52
problem. And by the way, this is probably
34:54
about the third one I've been through like this, right?
34:57
That's if you're in technology, guess what, prepare yourself.
35:03
So the the CEO Kevin had contacted me and I was kind of
35:10
the point of like and had I've been numerous contacts to, you
35:15
know, looking to fill is a role. Basically this one checked
35:21
several boxes for me. You know, the the healthcare,
35:25
I've kind of decided that's where I want to play, you know,
35:29
that the unfairness in our healthcare system, certainly in
35:33
the US speaks to me. It kind of like scratched
35:37
several witches. And that's why I kind of decided
35:39
to go with this. And aside from, yeah, OK, it's a
35:41
pretty interesting technical challenge as well.
35:44
And that that was kind of the trigger when I started thinking
35:47
about it was like, man, you know, there's data sets out
35:51
there that would be kind of cool to evaluate and apply against
35:54
this bias problem. And, you know, that's when
35:58
that's the one that kind of put me over.
35:59
I'm like OK yeah this is a real cool problem I solved too.
36:03
Well, what's fascinating also is that it's not like you have the
36:09
technical challenge of going out there and solving cancer.
36:14
It's like the technical challenge, you know, that if you
36:17
can just get people to take their pills or scan their heart
36:23
rate monitor or whatever it may be, these kind of low lifts for
36:28
people, then it can lead to a much longer life at a much
36:32
better outcome. And so the technical challenge
36:35
is more on the implementation of what is already in the world,
36:45
right? It's not like this cutting edge
36:47
bioscience or you have to go discover a new drug, any of
36:50
that. So that's also another piece
36:52
that I, I find fascinating. And yeah, now you get to
36:56
navigate more nuances of what kind of tone of voice are we
37:03
using to get people to do things.
37:05
And I imagine are you bringing in psychologist too to try and
37:09
help with this stuff or, or good copywriters, marketers that can
37:13
sell ice to an Eskimo. Yeah, I should say I.
37:17
Should mention we do have AI. Don't really know what her title
37:21
is, but she's a writer and she's advising on tone of voice.
37:27
And basically I gave her a whole set of question and answers from
37:32
our chat bot and she made a lot of Corrections on it.
37:35
So, so now I've got to go back and see if I can fix it right.
37:39
You know, that's the challenge is trying to figure out and, and
37:42
by the way, I, I say this before, but I think prompt
37:47
engineering is an oxymoron because to me there's, there's
37:51
are not a whole lot of engineering principles at at
37:53
least I can find to apply to it. You know, that's a great.
37:58
Point. It's throwing spaghetti at a
38:00
wall. Really.
38:02
Maybe I'm. Doing it wrong, I don't know.
38:04
But like, yeah. So it's, yeah, I mean, it's see,
38:08
you know, the interesting thing is it does seem like, you know,
38:12
technically it's, it should be an easy solve and then but
38:16
changing people's behaviour is really, really hard.
38:20
Yeah, you know, there you go. And and.
38:24
And kind of it brings my mind, one of the other things we're
38:27
looking into is, you know, all these wearable devices, using
38:31
those to monitor people remotely and then sending them a
38:35
notification and turning them into the chat bot to kind of
38:39
inform them about the, you know, the metrics that we see from
38:43
their monitoring device. This is actually something that
38:47
that I had done it with my team had done it Babylon.
38:52
We were looking at remote monitoring for blood pressure
38:57
and we had a stream of data that if we detected some metric out
39:05
of bound for for blood pressure, we would send a notification on
39:08
the app. And then that would put them
39:10
into it, a chat with our chat bot and inform them about, you
39:15
know, what it means and give them information and if they
39:20
would like to speak to a live agent to schedule an
39:22
appointment. So we're, we're actually looking
39:27
at something similar here. It's a Teo health as well, using
39:32
the remote monitoring device to and, you know, create a new
39:37
conversation so that they'll, you know, again, engagement in
39:41
the healthcare, right. Yeah.
39:43
I I wear a woop and I find it a little bit annoying when I get
39:52
the notifications from woop that will tell me how my sleep went.
39:56
And I also have a like 13 month old daughter and so my sleep
40:00
isn't the best. Actually my daughter wakes up
40:04
probably like 3-4 times a night still.
40:06
She's got really bad sleep and my wife takes the brunt of it.
40:10
She's the champion here, but I still am waking up every once in
40:14
a while and so. Whoop will send me these
40:18
notifications like your sleep, your sleep could have been
40:21
better last night. And it's like these passive
40:24
aggressive things like you might want to take a nap today.
40:27
It's like, man, I already feel like shit.
40:29
You're not helping the scenario here.
40:31
Like what's your deal, Whoop? You're giving me so many like
40:34
trying to just break me down. And so I feel like I am mentally
40:39
stronger because of all of the abuse that I've had it to take
40:44
from this whoop notifications over the last year.
40:47
Yeah, yeah. A good, a good friend of mine,
40:49
his mother has diabetes and he was monitoring.
40:52
And so he kept getting woken up at in the middle of night
40:55
because her glucose level was dropping below some threshold.
40:59
And so so he had to kind of figure out how to set her diet
41:03
up. But he was losing sleep because
41:04
his mom wasn't paying attention to her health.
41:07
So it's shit. Oh man.
41:09
So the these are the things now I just wanted to touch on this
41:14
evaluation again, because I think one thing that I've I've
41:18
thought about and I don't know if this is possible.
41:20
I imagine you've looked at it when it comes to guardrails and
41:25
it comes to, I know you mentioned you're using Ragas
41:27
right now. There's also other options out
41:31
there, like one called Guardrails AI and I know there's
41:36
Nemo Guardrails from NVIDIA. There's a tonne, right?
41:42
Have you not thought? About, hey, let's just throw
41:44
everything at it just in case. And so it can catch every single
41:50
edge case. Maybe Ragas is more, it's
41:52
stronger in this area. And then guardrails, we can set
41:55
up our specific guardrails we want or maybe I think they have
41:59
like templates for different verticals.
42:02
And then Nemo Guardrails has this strength?
42:05
Or is that just too redundant and overkill?
42:08
Well, I think so. I'll just kind of describe my
42:13
thinking and my thought process. So I'm using Ragas in a
42:20
development phase, right? So, so for example, I create a
42:26
baseline and when I of, of how this thing is behaving using
42:33
Ragas and so I'm logging those and MO flow.
42:36
So I have a history of those metrics, right?
42:40
And so I have a baseline and, and I'll just be specific.
42:45
I started out using Open AI ChatGPT 3.5.
42:53
I kind of did some kind of testing anecdotally with
42:56
different models, but this one just seemed easy and it was the
43:02
best for me at that time. So then I started doing this
43:07
study on hallucinations. So what I did now have a
43:12
baseline, right? And so I wanted to, I started
43:16
with the LLM. OK, so I ran the same tests
43:21
against literally every LLM I could find.
43:25
Well, I should say that's not pipe, that's a bit of hyperbole.
43:29
Every LLM in Amazon bedrock, right?
43:34
So now I have a baseline on a, you know, a dozen different
43:39
models and the metrics that I care about with regards to
43:45
hallucinations are the ones I'm paying attention to specifically
43:49
faithfulness, right? And so that helped me decide
43:52
which LLM to use. And it turns out, so that was I
43:56
think Anthropic Sonnet is the one that based on my test and I,
44:01
and I'll say this, don't trust me, trust it with your own data.
44:05
Everybody out there, this is not an easy pass.
44:08
You still have to do the hard work, you know, use it on your
44:12
data because it may behave differently depending on your
44:16
data, right? So that created a baseline based
44:22
on that model. Then the next thing I did was
44:26
testing against different embeddings models for the
44:30
retrieval and different of the parameters on the retrieval, you
44:35
know, the K values, etcetera. And So what I've been able to do
44:40
is create a baseline and incrementally move it based on
44:43
the metrics that I'm paying attention to, right?
44:47
So that's how I'm. Using Ragas, I'm also using it
44:52
to monitor and I'm not going to say, I'll say production, but I
44:58
we're not in production yet. We haven't delivered it in AVP
45:01
yet. We tend to do a, a pilot in
45:05
October for Mount Sinai Medical Centre with Mount Sinai Medical
45:11
Centre. So we'll start monitoring then,
45:15
right. We're probably not going to have
45:17
much throughput to really tell us a whole lot, but at least it
45:20
will start to give U.S. data. So I'm using Ragas to, you know,
45:25
measure against a baseline and and to monitor in place what I'm
45:31
so and then to take it to the next part of that discussion
45:35
about the guardrails. So I'm experimenting with
45:38
experimenting with Nemo guardrails.
45:41
That's NVIDIA open source project.
45:45
I'm looking at that for several things and and I'll tell you one
45:49
of the reasons why I like it. So the guardrails basically I'm
45:53
using it to test for prompt injection security violations to
45:59
make sure that chatbot doesn't say something stupid.
46:03
And so, you know, the guardrail will capture that.
46:05
And you know, I can either change the response or re, you
46:10
know, make some statement about like, sorry, don't I can't
46:15
respond to that, whatever like to be decided.
46:20
But what I also like about it is, is what something what we
46:24
had discussed earlier, I don't know if it was on or offline
46:30
about the need for something between intent based chat bots
46:36
and LLM based generative based chat bots.
46:39
Because in some cases I need to know what the response will be.
46:44
Right now, mind you, it's it's the intent is going to be
46:47
predicted, but at least I know what the response will be
46:51
because I'm writing it out myself or or whoever our product
46:55
people are clinical people are writing it out, you know how to
46:59
respond to specific intent. So that's how I intend to use
47:05
guardrails, not just kind of protect as guardrails, but also
47:10
for for an intent based type of response.
47:14
Yeah, it does feel like. Anytime anyone wants to know
47:18
information about other people and not themselves, that is a
47:24
but that's ripe for guardrails. But then it gets tricky again
47:29
because it's like, well, what are other people my age having
47:33
problems with? Those kind of questions can come
47:35
up, but if it's like, what is Eric having problems with?
47:39
Then the chat bot should be like, oh, well, Eric's fine and
47:43
you can ask him yourself. I'm not telling you anything
47:46
about Eric. But if it's like, Oh yeah,
47:48
people 40 to 45 generally are looking after these things or
47:54
these are the number one causes of health troubles in your area
47:59
or of people in your demographic with your weight high, that type
48:05
of thing, and your background, whatever it may be.
48:09
So yeah, it's a it's a fascinating problem to think
48:14
about. And I'm sure once you put it
48:17
into production, you're going to see.
48:20
All kinds of ways. That people are trying to mess
48:22
with it. Dude, I've already warned.
48:24
Everybody, man, like I've done all my testing.
48:28
We've got our ground truth that is like, you know, medically
48:32
certified. I guarantee you, you know,
48:35
haven't done this for a long, long time.
48:38
We put it out there with real people.
48:39
If you think it will go wrong, it certainly will help.
48:42
You know, like I've learned that the hard way.
48:45
Not now, mind you. You're like, we're not going to
48:47
have the throughput like it. You know, the Expedia we had, I
48:50
don't know, some of the, some of the models we were exposing to
48:54
like up to 2000 requests per second.
48:56
We're we're not going to have any if we have 2000 requests in
49:00
a week, cool. You know, like so you can kind
49:04
of babysit it. Especially in the beginning and
49:07
especially if you're working with one hospital or healthcare
49:11
centre, then it's really easy to be making sure that like, all
49:17
right, cool, this hasn't done what we didn't want it to do.
49:21
Yeah, yeah. We'll be monitoring.
49:23
And, and, and in fact, what, when we start the pilot, what we
49:26
intend to do is start with a subset, you know, start with 50
49:32
users, see how it behaves, see how they behave, see how they
49:36
react to it. They may hate it.
49:38
You know, we may have to start all over again.
49:40
Like who knows, right? I mean, that's, that's an
49:42
argument for, you know, getting exposure as early as possible.
49:47
Totally. You know, they may love it.
49:50
I'm I'm expecting to love it, right?
49:51
Like, yeah, open. Fingers crossed, yeah, but we'll
49:55
incrementally. You know, we'll learn from that,
49:57
you know, have several cycles what we'll learn, we'll, you
50:00
know, incrementally increase the traffic exposed and we'll learn
50:05
more each round of, you know, incrementally adding new users
50:08
to it, Right. Yeah.
50:10
Well, I think. It is awesome to hear and you're
50:14
adding so much value with a rag chat bot that I usually will
50:19
talk about how it adds no value to companies because you get the
50:24
HR experiences where it's like, great, now you spent all that
50:28
time, all that effort on that rag chat bot and your employees
50:32
know if they have 12 or 15 days of holidays that you could have
50:37
probably solved with search. But this proactive experience of
50:42
trying to get people to be more engaged with their healthcare
50:46
resonates with me. I love the vision and I thank
50:48
you for coming on here, Eric. Appreciate it.