WEBVTT
Kind: captions
Language: en-US

00:00:00.000 --> 00:00:00.500
[TEASER]

00:00:00.500 --> 00:00:01.000
[MUSIC PLAYS UNDER DIALOGUE]

00:00:01.000 --> 00:00:04.400
AKSHAY NAMBI: For me, research is just not about 
pushing the boundaries of the knowledge. It's  

00:00:04.400 --> 00:00:10.320
about ensuring that these advancements translate 
to meaningful impact on the ground. So, yes,  

00:00:10.320 --> 00:00:15.320
the big goals that guide most of my 
work is twofold. One, how do we build  

00:00:15.320 --> 00:00:20.320
technology that's scaled to benefit large 
populations? And two, at the same time,  

00:00:20.320 --> 00:00:27.880
I'm motivated by the challenge of tackling complex 
problems. That provides opportunity to explore,  

00:00:27.880 --> 00:00:33.167
learn, and also create something new, 
and that's what keeps me excited.

00:00:33.161 --> 00:00:34.027
[TEASER ENDS]

00:00:34.694 --> 00:00:37.160
CHRIS STETKIEWICZ: You're listening to Ideas, a 
Microsoft Research Podcast that dives deep into  

00:00:37.160 --> 00:00:50.600
the world of technology research and the profound 
questions behind the code. In this series, we'll  

00:00:50.600 --> 00:00:58.226
explore the technologies that are shaping our 
future and the big ideas that propel them forward.

00:00:58.226 --> 00:00:58.889
[MUSIC FADES]

00:00:58.889 --> 00:01:03.760
I'm your guest host, Chris Stetkiewicz. Today, 
I'm talking to Akshay Nambi. Akshay is a principal  

00:01:03.760 --> 00:01:09.400
researcher at Microsoft Research. His work lies 
at the intersection of systems, AI, and machine  

00:01:09.400 --> 00:01:14.280
learning with a focus on designing, deploying, 
and scaling AI systems to solve compelling  

00:01:14.280 --> 00:01:20.000
real-world problems. Akshay's research extends 
across education, agriculture, transportation,  

00:01:20.000 --> 00:01:25.440
and energy. He is currently working on enhancing 
the quality and reliability of AI systems by  

00:01:25.440 --> 00:01:30.840
addressing critical challenges such as reasoning, 
grounding, and managing complex queries.

00:01:30.840 --> 00:01:32.986
Akshay, welcome to the podcast.

00:01:32.986 --> 00:01:34.160
AKSHAY NAMBI: Thanks for having me.

00:01:34.160 --> 00:01:36.760
STETKIEWICZ: I'd like to begin by 
asking you to tell us your origin  

00:01:36.760 --> 00:01:41.200
story. How did you get started on 
your path? Was there a big idea or  

00:01:41.200 --> 00:01:45.760
experience that captured your imagination or 
motivated you to do what you're doing today?

00:01:45.760 --> 00:01:51.680
NAMBI: If I look back, my journey into research 
wasn't a straight line. It was more about  

00:01:51.680 --> 00:01:56.760
discovering my passion through some unexpected 
opportunities and also finding purpose along  

00:01:56.760 --> 00:02:02.720
the way. So before I started with my undergrad 
studies, I was very interested in electronics  

00:02:02.720 --> 00:02:07.120
and systems. My passion for electronics, kind 
of, started when I was in school. I was more  

00:02:07.120 --> 00:02:11.920
like an average student, not a nerd or not too 
curious, but I was always tinkering around,  

00:02:11.920 --> 00:02:17.600
doing things, building stuff, and playing with 
gadgets and that, kind of, made me very keen on  

00:02:17.600 --> 00:02:22.960
electronics and putting things together, and that 
was my passion. But sometimes things don't go as  

00:02:22.960 --> 00:02:28.160
planned. So I didn't get into the college which I 
had hoped to join for electronics, so I ended up  

00:02:28.160 --> 00:02:34.560
pursuing computer science, which wasn't too bad 
either. So during my final year of bachelor's,  

00:02:34.560 --> 00:02:39.640
I had to do a final semester project, which 
turned out to be a very pivotal moment. And  

00:02:39.640 --> 00:02:44.600
that's when I got to know this institute called 
Indian Institute of Science (IISc), which is a top  

00:02:44.600 --> 00:02:50.080
research institute in India and also globally. And 
I had a chance to work on a project there. And it  

00:02:50.080 --> 00:02:56.280
was my first real exposure to open-ended research, 
right, so I remember ... where we were trying to  

00:02:56.280 --> 00:03:02.400
build a solution that helped to efficiently 
construct an ontology for a specific domain,  

00:03:02.400 --> 00:03:06.880
which simply means that we were building 
systems to help users uncover relationships  

00:03:06.880 --> 00:03:12.040
in the data and allow them to query it more 
efficiently, right. And it was super exciting  

00:03:12.040 --> 00:03:17.800
for me to design and build something new. And that 
experience made me realize that I wanted to pursue  

00:03:17.800 --> 00:03:23.640
research further. And right after that project, 
I decided to explore research opportunities,  

00:03:23.640 --> 00:03:28.060
which led me to join Indian Institute of 
Science again as a research assistant.

00:03:28.060 --> 00:03:29.920
STETKIEWICZ: So what made you want to take the  

00:03:29.920 --> 00:03:32.920
skills you were developing and 
apply them to a research career?

00:03:32.920 --> 00:03:37.800
NAMBI: So interestingly when I joined IISc, 
the professor I worked with specialized in  

00:03:37.800 --> 00:03:43.160
electronics, so things come back, so something 
I had always been passionate about. And I was  

00:03:43.160 --> 00:03:48.440
the only computer science graduate in the lab at 
that time with others being electronic engineers,  

00:03:48.440 --> 00:03:54.400
and I didn't even know how to solder. But the lab 
environment was super encouraging, collaborative,  

00:03:54.400 --> 00:03:59.080
so I, kind of, caught up very quickly. In that 
lab, basically, I worked on several projects in  

00:03:59.080 --> 00:04:03.640
the emerging fields of embedded device and 
energy harvesting systems. Specifically,  

00:04:03.640 --> 00:04:09.080
we were designing systems that could harvest 
energy from sources like sun, hydro, and even  

00:04:09.080 --> 00:04:15.880
RF (radio frequency) signals. And my role was kind 
of twofold. One, I designed circuits and systems  

00:04:15.880 --> 00:04:21.680
to make energy harvesting more efficient so that 
you can store this energy. And then I also wrote  

00:04:21.680 --> 00:04:26.800
programs, software, to ensure that the harvested 
energy can be used efficiently. For instance,  

00:04:26.800 --> 00:04:31.320
as we harvest some of this energy, you want 
to have your programs run very quickly so that  

00:04:31.320 --> 00:04:36.960
you are able to sense the data, send it to the 
server in an efficient way. And one of the most  

00:04:36.960 --> 00:04:43.200
exciting projects I worked during that time was on 
data-driven agriculture. So this was back in 2008,  

00:04:43.200 --> 00:04:48.600
2009, right, where we developed an embedded system 
device with sensors to monitor the agricultural  

00:04:48.600 --> 00:04:54.080
fields, collecting data like soil moisture, soil 
temperature. And that was sent to the agronomists  

00:04:54.080 --> 00:04:59.760
who were able to analyze this data and provide 
feedback to farmers. In many remote areas, still  

00:04:59.760 --> 00:05:04.240
access to power is a huge challenge. So we used 
many of the technologies we were developing in the  

00:05:04.240 --> 00:05:08.920
lab, specifically energy harvesting techniques, 
to power these sensors and devices in the rural  

00:05:08.920 --> 00:05:14.720
farms, and that's when I really got to see 
firsthand how technology could help people's  

00:05:14.720 --> 00:05:21.040
lives, particularly in rural settings. And that's 
what, kind of, stood out in my experience at IISc,  

00:05:21.040 --> 00:05:26.600
right, was that it was [the] end-to-end nature 
of the work. And it was not just writing code or  

00:05:26.600 --> 00:05:32.480
designing circuits. It was about identifying the 
real-world problems, solving them efficiently,  

00:05:32.480 --> 00:05:37.280
and deploying solutions in the field. 
And this cemented my passion for creating  

00:05:37.280 --> 00:05:42.240
technology that solves real-world problems, 
and that's what keeps me driving even today.

00:05:42.240 --> 00:05:46.720
STETKIEWICZ: And as you're thinking about those 
problems that you want to try and solve, where did  

00:05:46.720 --> 00:05:50.940
you look for, for inspiration? It sounds like some 
of these are happening right there in your home.

00:05:50.940 --> 00:05:55.520
NAMBI: That's right. Growing up and living in 
India, I've been surrounded by these, kind of,  

00:05:55.520 --> 00:05:59.640
many challenges. And these are not distant 
problems. These are right in front of us.  

00:06:00.160 --> 00:06:05.200
And some of them are quite literally outside 
the door. So being here in India provides  

00:06:05.200 --> 00:06:10.520
a unique opportunity to tackle some of the 
pressing real-world challenges in agriculture,  

00:06:10.520 --> 00:06:16.980
education, or in road safety, where even small 
advancements can create significant impact.

00:06:16.980 --> 00:06:18.920
STETKIEWICZ: So how would you describe your  

00:06:18.920 --> 00:06:22.820
research philosophy? Do you have 
some big goals that guide you?

00:06:22.820 --> 00:06:28.400
NAMBI: Right, as I mentioned, right, my research 
philosophy is mainly rooted in solving real-world  

00:06:28.400 --> 00:06:33.920
problems through end-to-end innovation. For me, 
research is just not about pushing the boundaries  

00:06:33.920 --> 00:06:39.400
of the knowledge. It's about ensuring that these 
advancements translate to meaningful impact on  

00:06:39.400 --> 00:06:46.360
the ground, right. So, yes, the big goals that 
guide most of my work is twofold. One, how do  

00:06:46.360 --> 00:06:51.920
we build technology that's scaled to benefit 
large populations? And two, at the same time,  

00:06:51.920 --> 00:06:59.960
I'm motivated by the challenge of tackling complex 
problems. That provides opportunity to explore,  

00:06:59.960 --> 00:07:04.120
learn, and also create something new. 
And that's what keeps me excited.

00:07:04.120 --> 00:07:08.520
STETKIEWICZ: So let's talk a little bit about 
your journey at Microsoft Research. I know you  

00:07:08.520 --> 00:07:12.920
began as an intern, and some of the initial 
work you did was focused on computer vision,  

00:07:12.920 --> 00:07:16.800
road safety, energy efficiency. Tell 
us about some of those projects.

00:07:16.800 --> 00:07:22.160
NAMBI: As I was nearing the completion of my 
PhD, I was eager to look for opportunities  

00:07:22.160 --> 00:07:27.920
in industrial labs, and Microsoft Research 
obviously stood out as an exciting opportunity.  

00:07:27.920 --> 00:07:32.880
And additionally, the fact that Microsoft Research 
India was in my hometown, Bangalore, made it even  

00:07:32.880 --> 00:07:37.520
more appealing. So when I joined as an intern, 
I worked together with Venkat Padmanabhan,  

00:07:37.520 --> 00:07:42.320
who now leads the lab, and we started this 
project called HAMS, which stands for Harnessing  

00:07:42.320 --> 00:07:48.320
AutoMobiles for Safety. As you know, road 
safety is a major public health issue globally,  

00:07:48.320 --> 00:07:54.280
responsible for almost 1.35 million fatalities 
annually and with the situation being even more  

00:07:54.280 --> 00:07:59.560
severe in countries like India. For instance, 
there are estimates that there's a life lost  

00:07:59.560 --> 00:08:05.840
on the road every four minutes in India. When 
analyzing the factors which affect road safety,  

00:08:05.840 --> 00:08:10.800
we saw mainly three elements. One, the 
vehicle. Second, the infrastructure.  

00:08:10.800 --> 00:08:16.760
And then the driver. Among these, the driver 
plays the most critical role in many incidents,  

00:08:16.760 --> 00:08:22.160
whether it's over-speeding, driving without 
seat belts, drowsiness, fatigue, any of these,  

00:08:22.160 --> 00:08:28.920
right. And this realization motivated us to focus 
on driver monitoring, which led to the development  

00:08:28.920 --> 00:08:34.680
of HAMS. In a nutshell, HAMS is basically a 
smartphone-based system where you're mounting  

00:08:34.680 --> 00:08:40.000
your smartphone on a windshield of a vehicle 
to monitor both the driver and the driving in  

00:08:40.000 --> 00:08:45.800
real time with the goal of improving road safety. 
Basically, it observes key aspects such as where  

00:08:45.800 --> 00:08:50.640
the driver is looking, whether they are distracted 
or fatigued, while also considering the external  

00:08:50.640 --> 00:08:57.360
driving environment, because we truly believe 
to improve road safety, we need to understand  

00:08:57.360 --> 00:09:02.880
not just the driver's action but also the 
context in which they are driving. For example,  

00:09:02.880 --> 00:09:07.680
if the smartphone's accelerometer detects sharp 
braking, the system would automatically check the  

00:09:07.680 --> 00:09:11.720
distance to the vehicle in the front using 
the rear camera and whether the driver was  

00:09:11.720 --> 00:09:17.120
distracted or fatigued using the front camera. 
And this holistic approach ensures a more accurate  

00:09:17.120 --> 00:09:22.440
and comprehensive assessment of the driving 
behavior, enabling a more meaningful feedback.

00:09:22.440 --> 00:09:26.640
STETKIEWICZ: So that sounds like a system 
that's got several moving parts to it. And  

00:09:26.640 --> 00:09:30.920
I imagine you had some technical challenges you 
had to deal with there. Can you talk about that?

00:09:30.920 --> 00:09:36.880
NAMBI: One of our guiding principles in HAMS 
was to use commodity, off-the-shelf smartphone  

00:09:36.880 --> 00:09:42.480
devices, right. This should be affordable, in the 
range of $100 to $200, so that you can just take  

00:09:42.480 --> 00:09:48.400
out regular smartphones and enable this driver 
and driving monitoring. And that led to handling  

00:09:48.400 --> 00:09:53.560
several technical challenges. For instance, we had 
to develop efficient computer vision algorithms  

00:09:53.560 --> 00:09:59.720
that could run locally on the device with cheap 
smartphone processing units while still performing  

00:09:59.720 --> 00:10:04.800
very well at low-light conditions. We wrote 
multiple papers and developed many of the novel  

00:10:04.800 --> 00:10:10.680
algorithms which we implemented on very low-cost 
smartphones. And once we had such a monitoring  

00:10:10.680 --> 00:10:15.440
system, right, you can imagine there’s several 
deployment opportunities, starting from fleet  

00:10:15.440 --> 00:10:21.360
monitoring to even training new drivers, right. 
However, one application we hadn't originally  

00:10:21.360 --> 00:10:26.800
envisioned but turned out to be its most impactful 
use case even today is automated driver's  

00:10:26.800 --> 00:10:32.400
license testing. As you know, before you get a 
license, a driver is supposed to pass a test,  

00:10:32.400 --> 00:10:37.960
but what happens in many places, including 
India, is that licenses are issued with very  

00:10:37.960 --> 00:10:43.920
minimal or no actual testing, leading to unsafe 
and untrained drivers on the road. At the same  

00:10:43.920 --> 00:10:49.640
time as we were working on HAMS, Indian government 
were looking at introducing technology to make  

00:10:49.640 --> 00:10:55.280
testing more transparent and also automated. So 
we worked with the right set of partners, and we  

00:10:55.280 --> 00:10:59.760
demonstrated to the government that HAMS could 
actually completely automate the entire license  

00:10:59.760 --> 00:11:04.520
testing process. So we first deployed this system 
in Dehradun RTO (Regional Transport Office)—which  

00:11:04.520 --> 00:11:11.320
is the equivalent of a DMV in the US—in 2019, 
working very closely with RTO officials to define  

00:11:11.320 --> 00:11:16.120
what should be some of the evaluation criteria, 
right. Some of these would be very simple like,  

00:11:16.120 --> 00:11:20.320
oh, is it the same candidate who is taking 
the test who actually registered for the test,  

00:11:20.320 --> 00:11:24.080
right? And whether they are wearing seat belts. 
Did they scan their mirrors before taking a left  

00:11:24.080 --> 00:11:28.320
turn and how well they performed in tasks 
like reverse parking and things like that.

00:11:28.320 --> 00:11:30.800
STETKIEWICZ: So what's been the 
government response to that? Have  

00:11:30.800 --> 00:11:33.880
they embraced it or deployed it in a wider extent?

00:11:33.880 --> 00:11:40.480
NAMBI: Yes, yes. So after the deployment in 
Dehradun in 2019, we actually open sourced  

00:11:40.480 --> 00:11:45.320
the entire HAMS technology and our partners are 
now working with several state governments and  

00:11:45.320 --> 00:11:51.480
scaled HAMS to several states in India. And as 
of today, we have around 28 RTOs where HAMS is  

00:11:51.480 --> 00:11:57.640
actually being deployed. And the pass rate of 
such license test is just 60% as compared to  

00:11:57.640 --> 00:12:04.840
90-plus percent with manual testing. That's the 
extensive rigor the system brings in. And now what  

00:12:04.840 --> 00:12:10.240
excites me is after nearly five years later, we 
are now taking the next step in this project where  

00:12:10.240 --> 00:12:16.240
we are now evaluating the long-term impact of this 
intervention on driving behavior and road safety.  

00:12:16.240 --> 00:12:20.600
So we are collaborating with Professor Michael 
Kremer, who is a Nobel laureate and professor at  

00:12:20.600 --> 00:12:26.280
University of Chicago, and his team to study how 
this technology has influenced driving patterns  

00:12:26.280 --> 00:12:32.480
and accident rates over time. So this focus on 
closing the loop and moving beyond just deployment  

00:12:32.480 --> 00:12:36.880
in the field to actually measuring the real 
impact, right, is something that truly excites  

00:12:36.880 --> 00:12:42.280
me and that makes research at Microsoft is very 
unique. And that is actually one of the reasons  

00:12:42.280 --> 00:12:47.160
why I joined Microsoft Research as a full-time 
after my internship, and this unique flexibility  

00:12:47.160 --> 00:12:52.240
to work on real-world problems, develop novel 
research ideas, and actually collaborate with  

00:12:52.240 --> 00:12:58.300
partners both internally and externally to deploy 
at scale is something that is very unique here.

00:12:58.300 --> 00:13:00.160
STETKIEWICZ: So have you actually received  

00:13:00.160 --> 00:13:04.140
any evidence that the project is 
working? Is driving getting safer?

00:13:04.140 --> 00:13:09.080
NAMBI: Yes, these are very early analysis, and 
there are very positive insights we are getting  

00:13:09.080 --> 00:13:13.680
from that. Soon we will be releasing a white 
paper on our study on this long-term impact.

00:13:13.680 --> 00:13:18.080
STETKIEWICZ: That’s great. I look forward to 
that one. So you've also done some interesting  

00:13:18.080 --> 00:13:23.000
work involving the Internet of Things, with 
an emphasis on making it more reliable and  

00:13:23.000 --> 00:13:26.960
practical. So for those in our audience 
who may not know, the Internet of Things,  

00:13:26.960 --> 00:13:32.880
or IoT, is a network that includes billions 
of devices and sensors in things like smart  

00:13:32.880 --> 00:13:37.560
thermostats and fitness trackers. So talk 
a little bit about your work in this area.

00:13:37.560 --> 00:13:42.840
NAMBI: Right, so IoT, as you know, is already 
transforming several industries with billions  

00:13:42.840 --> 00:13:47.920
of sensors being deployed in areas like 
industrial monitoring, manufacturing,  

00:13:47.920 --> 00:13:53.160
agriculture, smart buildings, and also air 
pollution monitoring. And if you think about it,  

00:13:53.160 --> 00:13:58.720
these sensors provide critical data that 
businesses rely for decision making. However,  

00:13:58.720 --> 00:14:04.160
a fundamental challenge is ensuring that the 
data collected from these sensors is actually  

00:14:04.160 --> 00:14:10.640
reliable. If the data is faulty, it can lead 
to poor decisions and inefficiencies. And the  

00:14:10.640 --> 00:14:15.760
challenge is that these sensor failures are 
always not obvious. What I mean by that is  

00:14:15.760 --> 00:14:21.440
when a sensor stops working, it always doesn't 
stop sending data, but it often continues to  

00:14:21.440 --> 00:14:26.160
send some data which appear to be normal. 
And that's one of the biggest problems,  

00:14:26.160 --> 00:14:31.560
right. So detecting these errors is non-trivial 
because the faulty sensors can mimic real-world  

00:14:31.560 --> 00:14:36.840
working data, and traditional solutions like 
deploying redundant sensors or even manually  

00:14:36.840 --> 00:14:42.560
inspecting them are very expensive, labor 
intensive, and also sometimes infeasible,  

00:14:42.560 --> 00:14:48.960
especially for remote deployments. Our goal in 
this work was to develop a simple and efficient  

00:14:48.960 --> 00:14:54.480
way to remotely monitor the health of the IoT 
sensors. So what we did was we hypothesized  

00:14:54.480 --> 00:14:59.480
that most sensor failures occurred due to the 
electronic malfunctions. It could be either  

00:14:59.480 --> 00:15:04.960
due to short circuits or component degradation 
or due to environmental factors such as heat,  

00:15:04.960 --> 00:15:11.000
humidity, or pollution. Since these failures 
originate within the sensor hardware itself,  

00:15:11.000 --> 00:15:16.360
we saw an opportunity to leverage some of the 
basic electronic principles to create a novel  

00:15:16.360 --> 00:15:21.520
solution. The core idea was to develop a way 
to automatically generate a fingerprint for  

00:15:21.520 --> 00:15:26.960
each sensor. And by fingerprint, I mean the unique 
electrical characteristic exhibited by a properly  

00:15:26.960 --> 00:15:32.960
working sensor. We built a system that could 
devise these fingerprints for different types  

00:15:32.960 --> 00:15:39.040
of sensors, allowing us to detect failures purely 
based on the sensors internal characteristics,  

00:15:39.040 --> 00:15:44.320
that is the fingerprint, and even without looking 
at the data it produces. Essentially what it means  

00:15:44.320 --> 00:15:51.340
now is that we were able to tag each sensor data 
with a reliability score, ensuring verifiability.

00:15:51.340 --> 00:15:54.320
STETKIEWICZ: So how does that 
technology get deployed in the  

00:15:54.320 --> 00:15:57.720
real world? Is there an application 
where it's being put to work today?

00:15:57.720 --> 00:16:02.760
NAMBI: Yes, this technology, we worked 
together with Azure IoT and open-sourced it  

00:16:02.760 --> 00:16:08.200
where there were several opportunities and several 
companies took the solution into their systems,  

00:16:08.200 --> 00:16:13.280
including air pollution monitoring, smart 
buildings, industrial monitoring. The one  

00:16:13.280 --> 00:16:17.600
which I would like to talk about today is about 
air pollution monitoring. As you know, air  

00:16:17.600 --> 00:16:23.160
pollution is a major challenge in many parts of 
the world, especially in India. And traditionally,  

00:16:23.160 --> 00:16:29.360
air quality monitoring relies on these expensive 
fixed sensors, which provide limited coverage.  

00:16:29.360 --> 00:16:34.360
On the other hand, there is a rich body of 
work on low-cost sensors, which can offer  

00:16:34.360 --> 00:16:38.800
wider deployment. Like, you can put these sensors 
on a bus or a vehicle and have it move around  

00:16:38.800 --> 00:16:43.240
the entire city, where you can get much more 
fine-grained, accurate picture on the ground.  

00:16:43.240 --> 00:16:48.680
But these are often unreliable because these are 
low-cost sensors and have reliability issues. So  

00:16:48.680 --> 00:16:53.040
we collaborated with several startups who were 
developing these low-cost air pollution sensors  

00:16:53.040 --> 00:16:57.760
who were finding it very challenging to gain 
trust because one of the main concerns was the  

00:16:57.760 --> 00:17:03.480
accuracy of the data from low-cost sensors. So our 
solution seamlessly integrated with these sensors,  

00:17:03.480 --> 00:17:07.320
which enabled verification of the data 
quality coming out from these low-cost  

00:17:07.320 --> 00:17:12.360
air pollution sensors. So this bridged the 
trust gap, allowing government agencies  

00:17:12.360 --> 00:17:17.420
to initiate large-scale pilots using low-cost 
sensors for fine-grain air-quality monitoring.

00:17:17.420 --> 00:17:21.520
STETKIEWICZ: So as we're talking about 
evolving technology, large language models,  

00:17:21.520 --> 00:17:26.480
or LLMs, are also enabling big changes, and 
they're not theoretical. They're happening  

00:17:26.480 --> 00:17:30.920
today. And you've been working on LLMs 
and their applicability to real-world  

00:17:30.920 --> 00:17:34.480
problems. Can you talk about your work 
there and some of the latest releases?

00:17:34.480 --> 00:17:41.040
NAMBI: So when ChatGPT was first released, I, 
like many people, was very skeptical. However,  

00:17:41.040 --> 00:17:46.600
I was also curious both of how it worked and, 
more importantly, whether it could accelerate  

00:17:46.600 --> 00:17:52.640
solutions to real-world problems. That led to 
the exploration of LLMs in education, where  

00:17:52.640 --> 00:17:57.840
we fundamentally asked this question, can AI help 
improve educational outcomes? And this was one of  

00:17:57.840 --> 00:18:03.320
the key questions which led to the development 
of Shiksha copilot, which is a genAI-powered  

00:18:03.320 --> 00:18:08.280
assistant designed to support teachers in their 
daily work, starting from helping them to create  

00:18:08.280 --> 00:18:12.640
personalized learning experience, design 
assignments, generate hands-on activities,  

00:18:12.640 --> 00:18:18.240
and even more. Teachers today universally face 
several challenges, from time management to lesson  

00:18:18.240 --> 00:18:24.040
planning. And our goal with Shiksha was to empower 
them to significantly reduce the time spent on  

00:18:24.040 --> 00:18:28.920
this task. For instance, lesson planning, 
which traditionally took about 60 minutes,  

00:18:28.920 --> 00:18:34.000
can now be completed in just five minutes using 
the Shiksha copilot. And what makes Shiksha unique  

00:18:34.000 --> 00:18:39.200
is that it's completely grounded in the local 
curriculum and the learning objectives, ensuring  

00:18:39.200 --> 00:18:44.640
that the AI-generated content aligns very well 
with the pedagogical best practices. The system  

00:18:44.640 --> 00:18:49.600
actually supports multilingual interactions, 
multimodal capabilities, and also integration  

00:18:49.600 --> 00:18:54.240
with external knowledge base, making it very 
highly adaptable for different curriculums.  

00:18:54.240 --> 00:18:58.720
Initially, many teachers were skeptical. Some 
feared this would limit their creativity.  

00:18:59.600 --> 00:19:04.000
However, as they began starting to use Shiksha, 
they realized that it didn't replace their  

00:19:04.000 --> 00:19:09.520
expertise, but rather amplified it, enabling 
them to do work faster and more efficiently.

00:19:09.520 --> 00:19:13.360
STETKIEWICZ: So, Akshay, the last time 
you and I talked about Shiksha copilot,  

00:19:13.360 --> 00:19:17.600
it was very much in the pilot phase and the 
teachers were just getting their hands on it.  

00:19:17.600 --> 00:19:21.060
So it sounds like, though, you've gotten some 
pretty good feedback from them since then.

00:19:21.060 --> 00:19:25.200
NAMBI: Yes, so when we were discussing, we 
were doing this six-month pilot with 50-plus  

00:19:25.200 --> 00:19:30.720
teachers where we gathered overwhelming positive 
feedback on how technologies are helping teachers  

00:19:30.720 --> 00:19:34.720
to reduce time in their lesson planning. 
And in fact, they were using the system so  

00:19:34.720 --> 00:19:39.440
much that they really enjoyed working with 
Shiksha copilot where they were able to do  

00:19:39.440 --> 00:19:43.960
more things with much less time, right. 
And with a lot of feedback from teachers,  

00:19:43.960 --> 00:19:49.480
we have improved Shiksha copilot over the past few 
months. And starting this academic year, we have  

00:19:49.480 --> 00:19:54.480
already deployed Shiksha to 1,000-plus teachers in 
Karnataka. This is with close collaboration with  

00:19:54.480 --> 00:19:58.640
our partners in … with the Sikshana Foundation 
and also with the government of Karnataka.  

00:19:59.280 --> 00:20:04.200
And the response has been already incredibly 
encouraging. And looking ahead, we are actually  

00:20:04.200 --> 00:20:09.000
focusing on again, closing this loop, right, and 
measuring the impact on the ground, where we are  

00:20:09.000 --> 00:20:13.760
doing a lot of studies with the teachers to 
understand not just improving efficiency of  

00:20:13.760 --> 00:20:20.160
the teachers but also measuring how AI-generated 
content enriched by teachers is actually enhancing  

00:20:20.160 --> 00:20:24.560
student learning objectives. So that's the study 
we are conducting, which hopefully will close this  

00:20:24.560 --> 00:20:30.020
loop and understand our original question that, 
can AI actually help improve educational outcomes?

00:20:30.020 --> 00:20:34.560
STETKIEWICZ: And is the deployment 
primarily in rural areas, or does  

00:20:34.560 --> 00:20:37.900
it include urban centers, or what's the target?

00:20:37.900 --> 00:20:41.760
NAMBI: So the current deployment with 
1,000 teachers is a combination of  

00:20:41.760 --> 00:20:46.040
both rural and urban public schools. 
These are covering both English medium  

00:20:46.040 --> 00:20:51.180
and Kannada medium teaching schools 
with grades from Class 5 to Class 10.

00:20:51.180 --> 00:20:56.320
STETKIEWICZ: Great. So Shiksha was focused on 
helping teachers and making their jobs easier,  

00:20:56.320 --> 00:20:58.880
but I understand you're also 
working on some opportunities  

00:20:58.880 --> 00:21:02.420
to use AI to help students 
succeed. Can you talk about that?

00:21:02.420 --> 00:21:08.480
NAMBI: So as you know, LLMs are still evolving and 
inherently they are fragile, and deploying them in  

00:21:08.480 --> 00:21:13.360
real-world settings, especially in education, 
presents a lot of challenges. With Shiksha,  

00:21:13.360 --> 00:21:17.600
if you think about it, teachers remain 
in control throughout the interaction,  

00:21:17.600 --> 00:21:21.440
making the final decision on whether to use 
the AI-generated content in the classroom  

00:21:21.440 --> 00:21:26.960
or not. However, when it comes to AI tutors 
for students, the stakes are slightly higher,  

00:21:26.960 --> 00:21:32.240
where we need to ensure the AI doesn't produce 
incorrect answers, misrepresent concepts,  

00:21:32.240 --> 00:21:38.200
or even mislead explanations. Currently, we are 
developing solutions to enhance accuracy and also  

00:21:38.200 --> 00:21:43.880
the reasoning capabilities of these foundational 
models, particularly solving math problems.  

00:21:43.880 --> 00:21:49.920
This represents a major step towards building AI 
systems that's much more holistic personal tutors,  

00:21:49.920 --> 00:21:54.700
which help student understanding and create 
more engaging, effective learning experience.

00:21:54.700 --> 00:22:00.960
STETKIEWICZ: So you've talked about working 
in computer vision and IoT and LLMs. What do  

00:22:00.960 --> 00:22:05.000
those areas have in common? Is there some thread 
that weaves through the work that you're doing?

00:22:05.000 --> 00:22:09.000
NAMBI: That's a great question. As a systems 
researcher, I'm quite interested in this  

00:22:09.000 --> 00:22:14.280
end-to-end systems development, which means 
that my focus is not just about improving a  

00:22:14.280 --> 00:22:19.280
particular algorithm but also thinking about 
the end-to-end system, which means that I,  

00:22:19.280 --> 00:22:23.960
kind of, think about computer vision, 
IoT, and even LLMs as tools, where we  

00:22:23.960 --> 00:22:28.720
would want to improve them for a particular 
application. It could be agriculture, education,  

00:22:28.720 --> 00:22:34.040
or road safety. And then how do you think this 
holistically to come up with the best efficient  

00:22:34.040 --> 00:22:39.320
system that can be deployed at population scale, 
right. I think that's the connecting story here,  

00:22:39.320 --> 00:22:44.200
that how do you have this systemic thinking 
which kind of takes the existing tools,  

00:22:44.200 --> 00:22:49.460
improves them, makes it more efficient, and 
takes it out from the lab to real world.

00:22:49.460 --> 00:22:53.640
STETKIEWICZ: So you're working on some 
very powerful technology that is creating  

00:22:53.640 --> 00:22:58.600
tangible benefits for society, which is your 
goal. At the same time, we're still in the  

00:22:58.600 --> 00:23:03.320
very early stages of the development of AI and 
machine learning. Have you ever thought about  

00:23:03.320 --> 00:23:07.560
unintended consequences? Are there some 
things that could go wrong, even if we  

00:23:07.560 --> 00:23:12.620
get the technology right? And does that kind of 
thinking ever influence the development process?

00:23:12.620 --> 00:23:17.800
NAMBI: Absolutely. Unintended consequences 
are something I think about deeply. Even  

00:23:17.800 --> 00:23:22.600
the most well-designed technology can have these 
ripple effects that we may not fully anticipate,  

00:23:22.600 --> 00:23:27.120
especially when we are deploying it at 
population scale. For me, being proactive  

00:23:27.120 --> 00:23:31.120
is one of the key important aspects. This 
means not only designing the technology at  

00:23:31.120 --> 00:23:36.080
the lab but actually also carefully deploying 
them in real world, measuring its impact,  

00:23:36.080 --> 00:23:40.400
and working with the stakeholders to minimize 
the harm. In most of my work, I try to work  

00:23:40.400 --> 00:23:45.360
very closely with the partner team on the ground 
to monitor, analyze, how the technology is being  

00:23:45.360 --> 00:23:50.360
used and what are some of the risks and how can 
we eliminate that. At the same time, I also remain  

00:23:50.360 --> 00:23:56.000
very optimistic. It's also about responsibility. 
If we are able to embed societal values, ethics,  

00:23:56.000 --> 00:24:00.200
into the design of the system and involve 
diverse perspectives, especially from people  

00:24:00.200 --> 00:24:05.560
on the ground, we can remain vigilant as the 
technology evolves and we can create systems  

00:24:05.560 --> 00:24:10.480
that can truly deliver immense societal benefits 
while addressing many of the potential risks.

00:24:10.480 --> 00:24:17.000
STETKIEWICZ: So we've heard a lot of great 
examples today about building technology to  

00:24:17.000 --> 00:24:21.560
solve real-world problems and your motivation 
to keep doing that. So as you look ahead,  

00:24:21.560 --> 00:24:25.920
where do you see your research going 
next? How will people be better off  

00:24:25.920 --> 00:24:29.740
because of the technology you develop 
and the advances that they support?

00:24:29.740 --> 00:24:36.200
NAMBI: Yeah, I'm deeply interested in advancing 
AI systems that can truly assist anyone in their  

00:24:36.200 --> 00:24:41.320
daily tasks, whether it's providing personalized 
guidance to a farmer in a rural village, helping  

00:24:41.320 --> 00:24:46.360
a student get instant 24 by 7 support for their 
learning doubts, or even empowering professionals  

00:24:46.360 --> 00:24:51.880
to work more efficiently. And to achieve this, 
my research is focusing on tackling some of the  

00:24:51.880 --> 00:24:57.240
fundamental challenges in AI with respect to 
reasoning and reliability and also making sure  

00:24:57.240 --> 00:25:03.120
that AI is more context aware and responsive 
to evolving user needs. And looking ahead,  

00:25:03.120 --> 00:25:08.720
I envision AI as not just an assistant but 
also as an intelligent and equitable copilot  

00:25:08.720 --> 00:25:14.140
seamlessly integrated into our everyday life, 
empowering individuals across various domains.

00:25:14.140 --> 00:25:20.105
STETKIEWICZ: Great. Well, Akshay, thank you 
for joining us on Ideas. It's been a pleasure.

00:25:20.105 --> 00:25:20.766
[MUSIC]

00:25:20.766 --> 00:25:22.680
NAMBI: Yeah, I really enjoyed 
talking to you, Chris. Thank you.

00:25:22.680 --> 00:25:28.320
STETKIEWICZ: Till next time.

00:25:28.320 --> 00:25:29.126
[MUSIC FADES]