01
NEWSLETTER
JAN 15 - JULY 15, 2026
VOL 04 - ISSUE 01
Christened ‘1729', after Ramanujan - Hardy number, the Plaksha
newsletter is a window into our thriving, interconnected and
learner centered environment where Plakshans look beyond the
obvious, just like Srinivasa Ramanujan did with the seemingly dull
number ‘1729'.
I
ndia's problems are, in many ways, different from the world's problems, in nature
and in scale. Moreover, nearly two of every three Indians are under 35. Ten to
twelve million Indians enter the workforce every year, which is in stark contrast to
the developed world’s shrinking workforce. A conversation about technology, specifi-
cally AI, for India must be in this context.
Can AI help solve the problems of a farmer who loses a season's income because
information about a pest outbreak reaches him too late, or can it solve the woes of a
city whose bus routes were decided before a model of how people actually move
through it was created? The answer is yes, but the solutions made for the developed
world have to be India-specific, created by experts who know the domain and the
technology, committed to the work for years not quarters.
Such solutions seldom come from companies or government departments. Companies
respond to market signals, market around most of these problems has not formed yet.
Governments often answer to an election cycle, but building AI for India will outlast
several of them. Universities are among the few institutions built to absorb both: to
fund research long before a market exists, and to train, one cohort at a time, people
who will carry that research into practice over careers, not quarters.
Plaksha’s commitment to building India’s AI engine is reflected in the interdisciplinary
centers and in the researchers. Our interdisciplinary research centers in agriculture,
clean energy, health and water are using AI to address grand challenges. We are also
nurturing tech leaders, researchers and entrepreneurs who are equally fluent in the
domain and the technology, and giving them the time, in years, to do the work properly.
This newsletter gives a peek into these efforts.
Rudra Pratap
Founding Vice Chancellor
Plaksha University
BUILDING INDIA’S AI ENGINE
T H I N K
Alignment is the Key to Safe AI
When multiple AI agents interact at machine speed, closing the gap between their individual utility and the societal
good becomes the real test of safety.
S
ome months ago, I ran a retreat with cap-
tains of industry in a beautiful place on the
California coast called Bodega Bay, where
we discussed technology competitiveness. A few
themes emerged clearly. One is the perception that
AI will change everything. Another is the revival of
semiconductor manufacturing because of a surge
in the demand for both processors and memory
chips to power the appetite for computation. We,
the consumers, are already paying the price in the
form of higher PC and smartphone prices. At the
same time, new advances in quantum sensing and
communications are galloping along. While there is
no dominant, scalable quantum technology yet, the
search for new computational substrates is driving
several new technologies.
Furthermore, physical AI, which I refer to as em-
bodied intelligence, is going to drive edge intelli-
gence and a new set of power-lean CPU+GPU pro-
cessors. For instance, in a new laboratory that I
have launched called EMBER, for Embodied
Intelligence and Robotics, we are hard at work de-
veloping humanoids, multi-fingered hands, and
even a robotic Formula 1–like IndyCar. Our human-
oid robots can assemble other robots, perform
search and rescue, and even play table tennis.
LLMs and VLAs (Vision-Language-Action models)
are great in terms of translating natural language
descriptions into code, but natural language is a
coarse abstraction of detailed motor actions. Thus,
we have a lot further to go to keep them from hal-
lucinating.
LLMs and foundation models are changing how we
live, but natural language is not an adequate ab-
straction for machines interacting with the environ-
ment. The 2025 Turing Award winner and father of
reinforcement learning, Richard Sutton, in his fa-
mous essay called ‘The Bitter Lesson’, posited that
having more data will enable computers to outper-
form all analytical approaches. This seems to indi-
cate that we can dispense with the need to learn
about model-based approaches for the design of in-
telligent systems, sometimes referred to as
neurosymbolic reasoning. However, I fear that an
even more bitter lesson may be that investing yet
more money to collect more data and build more
GPU farms may break the bank first!
This brings me to a real important part of this ad-
dress: safe AI. While it is true that Claude, Gemini,
ChatGPT, Qwen and Kimi are generating fantastic
agentic code and are certainly giving programmers
a run for their money, builders of large systems are
increasingly worried about how to provide assur-
ances of correctness for the agentic models.
Furthermore, when multiple agentic models inter-
act with each other at machine speed, we need to
be able to provide alignment between the agents.
(Edited excerpt from Prof Sastry’s Convocation Address to
Plaksha’s Undergraduate Class of 2026)
S. Shankar Sastry
Founding Chancellor
Plaksha University
Alignment is a key ingredient of safe AI.
Computational game theory gives us a way of
defining a so-called Nash equilibrium as a
state in which non-cooperative agents play to
maximize their utilities in a no-regret equilib-
rium. The problem with these Nash equilibria
is that there is a gap between the collective
utility of these players and the societal com-
mon good. This is referred to as the ‘price of
anarchy’ or the ‘tragedy of the commons.’ We
see this each day that we drive. Our desire
for a Nash equilibrium in minimizing our own
transit time is frequently at odds with societal
goals, such as reducing the average transit
time or improving the utilization of the roads.
By alignment of multi-agent LLMs is meant
the minimization of the price of anarchy
through incentives and societal mechanism
design. It is not yet settled what form of soci-
etal guardrails is needed to provide this kind
of safe AI while not stifling innovation.
02
Autobiography of AI
1950
The Early Years
My story began when Alan Turing published
‘Computing Machinery and Intelligence’ and asked
whether machines could think. His proposal
became the Turing Test, used to evaluate if a
machine could imitate human conversation. At
the same time, the first rule-based generative
systems emerged: programs that could produce
outputs by following expert-designed rules.
BUILDING INDIA’S AI ENGINE
Building India’s AI Engine | AI for Humanity | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure | Developing Algorithms for
Economic Value | Research at Plaksha | Student Innovations
BUILDING INDIA’S AI ENGINE
Shaping the Future
of AI for Humanity
The next leap in AI will be about building systems
with fairness, accountability and human context.
I
ndia’s AI Impact Summit was based on the
theme ‘Sarvajana Hitaya, Sarvajana
Sukhaya’, which means welfare and well-
being for all. It reinforced a timely idea: the fu-
ture of AI must be guided by AI for Humanity.
Achieving this vision, however, depends not just
on technological progress, but on how con-
sciously we design systems to be inclusive and
accessible.
Inclusivity requires AI systems to be fair and un-
biased in their outcomes, while democratization
ensures that access to AI and the ability to
shape it is widely distributed. Without these, AI
risks reinforcing the very inequalities it promises
to solve.
The consequences of neglecting fairness are so
far visible. Variables like zip codes, often used in
lending, insurance, or hiring, act as proxies for
socio-economic disparities. Similarly, widely dis-
cussed cases, from biased recidivism prediction
systems to gender-skewed recruitment tools
highlight how AI misclassifies or places certain
groups at a disadvantage despite appearing
technically robust. These examples make one
point clear: accuracy alone cannot define a re-
sponsible AI system.
Addressing this issue requires a lifecycle per-
spective. Ethical considerations ought to begin
at problem formulation - what is being opti-
mized and for whom and continue through
data collection, annotation, model development
and deployment. Given that bias can enter at
any stage, consistent evaluation and monitoring
are essential.
Responsibility in AI is collective. Educational in-
stitutions should equip students not only with
technical skills but also ethical reasoning.
Industry must embed accountability into design
and deployment practices. Policymakers play a
critical role in setting guardrails; regulatory
efforts such as the European Union’s AI Act
mark important steps toward governing AI re-
sponsibly.
As AI systems become deeply embedded in so-
cially impactful domains, the challenge is no
longer just building powerful models but
building them responsibly. The path forward
lies in aligning innovation with intent, ensuring
AI advances not only efficiency and scale, but
also fairness, dignity and shared societal prog-
ress.
Rajesh Sharma
Associate Dean
Harish and Bina Shah School of AI and
Computer Science,
Plaksha University
03
1956
My Birth at Dartmouth
In 1956, scientists worked on the Dartmouth
Summer Research Project on AI. It was here that
the academic discipline of AI was officially
launched. This was the beginning of computer
programs that could solve puzzles, prove
mathematical theorems and perform symbolic
reasoning.
T H I N K
Building India’s AI Engine | Alignment is the Key to Safe AI | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure | Developing Algorithms
for Economic Value | Research at Plaksha | Student Innovations
Two AI agents on a phone call figure out neither is human, then switch
from English to data-over-sound.
BUILDING INDIA’S AI ENGINE
AI as a Tool to Question
As chatbots enter classrooms, educators face a big challenge:
teaching students to direct AI, rather than accept its answers
without scrutiny.
T
here is a particular mood that falls over a
classroom when someone understands
something for the first time. It is the quiet
sound of attention gathering and an idea being
formed. As both a professor and a learner, I love
and covet these rewarding moments. In a class-
room, students share their ideas, discuss, argue
and challenge each other to think more deeply and
expand their thinking in directions they might not
have ever considered. If I am alone and outside of
the classroom, I might go down an information rab-
bit hole, where I read as much as I can about a
topic, and talk to a friend who might be equally ex-
cited about it or teach me something I might not
have known. Learning is an experience of building
something inside yourself (or with others) that was
not there before. It is one of the oldest human ex-
periences, and one that will exist in the future. It is
also notoriously hard to study, as learning is full of
social complexities, relationships and human inter-
actions, which tend to be the opposite of codified
interactions where quantitative evaluation will re-
make an intervention into its own reflection.
Unlike previous classroom technologies, AI did not
go through the usual process of a school deliber-
ately deciding to adopt it. Researchers who study
education technology call this pattern of adoption
an arrival technology: where a technology enters
everyday life and alters environments spontane-
ously, thereby forcing a reaction. Decades of re-
search on education technology show how new
tools rarely transform education; they get folded
into what a system already values and tend to ben-
efit the students who were already best posi-
tioned to succeed. This illustrates how AI (or any
technology) will not correct educational chal-
lenges on its own. Researchers who study learner
agency know that classrooms built around stu-
dents creating things that matter to them, rather
than just receiving information, create experi-
ences that students carry with them beyond sim-
ply performing on a test. In that kind of class-
room, AI becomes a tool that students direct and
question, not a tool or answer they blindly accept.
This kind of learning and thinking with AI and tech-
nologies is not foreign to faculty and students at
Plaksha. Instead, it is central to teaching and learn-
ing at Plaksha. Students learn how to build real
things from and with faculty: robots, research pro-
jects, solutions to real problems. Or in my own
class, where I teach Engineering Ethics, we read,
discuss and write about the impacts of technolo-
gies, in both creative and critical ways, because the
inquiries that students have about technologies
and their impacts matter as much as the technol-
ogy. Students at Plaksha learn to define and ad-
dress issues from a multitude of different perspec-
tives, interests and disciplines, which strengthens
their understanding of what it means to create
meaningful impact in the world.
Jenny Tilsen
Assistant Professor
Plaksha University
04
1957
The First Neural Network
The Perceptron, one of the first artificial neural
networks, enabled me to pick learning patterns
from data. Through it, I could learn for the first
time. The mathematical roots of this moment
traced back to feed-forward neural networks from
the 1950s and the foundational work on neural
computation by McCulloch and Pitts in 1943.
Created by Frank Rosenblatt, it seeded deep
learning models, transformers and large language
models. But the Perceptron had a limitation. It
could only solve linearly separable problems.
When it was exposed in 1969, funding dried up.
T H I N K
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure | Developing Algorithms for
Economic Value | Research at Plaksha | Student Innovations
BUILDING INDIA’S AI ENGINE
B U I L D
Establishing Ecosystems for Building AI
People, partnerships and platforms that take AI from possibility to impact.
Building Global AI Alliances
Researchers, educators, policymakers, and
industry leaders from across the world discussed
ways to integrate pedagogy, infrastructure and
real-world application of AI into a values-driven
system of impact at the US–India Research
Workshop. The depth of engagement, ranging
from high-level deliberation to nuanced debates
on AI hallucinations, bias, governance and human-
AI collaboration stood out. The lesson running
through all of it was that responsible AI is a design
question, settled in how systems are taught, built
and evaluated.
Speakers and key dignitaries at the US–India AI
Research Workshop.
05
1966
I Learned to Converse
ELIZA, one of my first chatbot forms, was
introduced at MIT. ELIZA acted as a
psychotherapist, responding to patient inputs
through pattern matching and a library of pre-
designed rule-based responses. But ELIZA
could not understand context, remember past
exchanges or reason across topics.
B U I L D
Grounding AI Research in Indian Realities
At ConfAI 2025, researchers examined what it will take to make AI systems trustworthy and reliable.
AI systems built elsewhere carry assumptions that
do not survive contact with Indian conditions,
whether that is a clinic without specialists, a city
utility under strain or a lending decision made on
thin data. Work on those problems is scattered
across institutions that rarely share a program.
ConfAI brings that work onto a single platform,
and gives students access to researchers they
would otherwise only read about.
It features seven tracks: AI in the Economy; AI in
Healthcare; AI in Sensors; Reasoning in Vision
Language Models; AI4Cities; Responsible and
Ethical AI; and AI and Logic. The annual
conference on artificial intelligence is hosted by
the Harish and Bina Shah School of AI and
Computer Science at Plaksha.
A typical edition draws more than 600 students,
over 100 visitors, five or more keynote speakers
and around 20 oral and 20 poster presentations.
Tracks are chaired by Plaksha faculty, which keeps
the program tied to live research rather than to a
fixed thematic list.
India’s AI engine will not be built by algorithms only; it will be shaped by global collaborations that strengthen local
innovation, entrepreneurs, who transform breakthroughs into companies, and industry that helps prepare the talent that
will define the future. At Plaksha, this ecosystem is taking shape through a series of initiatives that span international
research partnerships, startup incubation, inclusive innovation, and industry-led learning.
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Infrastructure | Developing Algorithms for Economic Value |
Research at Plaksha | Student Innovations
BUILDING INDIA’S AI ENGINE
Strengthening Public Digital Infrastructure
Through AI
Plaksha's AI and data science teams are developing AI-enabled tools for
Punjab government’s portal, building an AI layer on top of existing approval
systems. The tools are designed to simplify approvals, reduce friction for
MSMEs and make regulatory processes more transparent, predictable and
efficient for businesses.
Fast Track Portal, Punjab’s single-window clearance platform is set to gain
new AI capabilities through a collaboration with Plaksha University.
Shaping an
Inclusive AI Future
SheBuildsAI hosted founders, funders, researchers
and government representatives to answer a
question that sits at the foundation of AI
development - who builds it and whose reality it
learns from. The stakes are high. Women make up
22% of the global AI workforce and receive under
2% of global VC funding. Left unaddressed, this
imbalance gets built into the systems AI produces.
SheBuildsAI saw founder-investor funding,
government agency opening, discussions on
gender-disaggregated data collection frameworks,
and a women's leadership network connecting
founders, researchers and funders.
(L-R): Shaili Chopra, Tanuja Ganu, Soumya Batra and Sukhman Randhawa, in a panel discussion moderated by
Sarika Bhattacharyya.
06
1970
First AI Winter
Though the 1970s brought a long season of
reduced investment, statistical machine learning
started taking shape, with probabilistic graphical
models, Markov chains, and the expectation-
maximization algorithm offering new ways to learn
from data.
Researchers began building rigorous foundations
in mathematics, statistics and probability theory.
But I lacked the computational power, data and
the algorithms needed to scale.
B U I L D
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Infrastructure | Developing Algorithms for Economic Value |
Research at Plaksha | Student Innovations
BUILDING INDIA’S AI ENGINE
B U I L D
Infrastructure Powering Innovation
By integrating intelligent systems, sustainable design and advanced research facilities, Plaksha is building a
campus equipped to support the computing demands of modern AI research.
D
espite being built during COVID-19, a
time of uncertainty, constrained re-
sources and constant ambiguity,
Plaksha’s infrastructure is intentional and was de-
signed to ensure what we create will remain rele-
vant well into the future. The reference point was
never local; it was what researchers anywhere in
the world would need to do serious work in AI. So,
the university set out to create an ecosystem
where education, research and industry could
come together to accelerate innovation. That vi-
sion continues to shape a campus on par with lead-
ing global institutions.
Each iteration at the development stage made our
classrooms and labs more responsive to the needs
of learners and educators. A key outcome of this is
the Center for Thinking, Language and
Communication, dedicated to bringing human-
centered communication and expression to the
forefront of engineering education.
Our early efforts also led to the establishment of
advanced research labs equipped with high-
performance computing, collaborative workspaces
and flexible environments that enable experimen-
tation. Complementing this, our robotics center
serves as a space where theory meets real-world
application.
Across the campus, AI-enabled platforms person-
alize learning, support faculty with real-time in-
sights and make education more engaging. In hos-
tels and shared facilities, tools such as Building
Management Systems (BMS), smart access con-
trols and recognition technologies streamline oper-
ations and improve efficiency. Smart HVAC optimi-
zation and regular energy audits keep buildings
efficient without compromising comfort, while
thoughtful data center design and predictive facili-
ties management ensure the infrastructure be-
hind our research remains reliable as computing
needs continue to grow. Designed to work quietly
in the background, these systems allow the uni-
versity to focus on its core mission of learning and
discovery.
Jitin Sahni
Chief Operating Officer
Plaksha University
07
1980
Expert Systems Revive Me
Two parallel streams defined this decade. First, expert
systems, encoding specialist knowledge into rule-
based programs. Second, statistical machine
translation which replaced rule-based translation.
Autoregressive models for language generation also
appeared in this era.
Businesses began investing in me. I proved to be
useful in specialist domains: medical diagnosis,
engineering configuration and financial analysis.
Nevertheless, expert systems were expensive, brittle
and impossible to maintain at scale. They could not
adapt to new situations. Statistical models were more
flexible but limited by dimensionality and
computational resources.
B U I L D
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | Developing Algorithms
for Economic Value | Research at Plaksha | Student Innovations
The Harish and Bina Shah School of AI and
Computer Science
From code to motion: The ARK Robotics LabThe Dixon IoT Lab, set up for hands-on learning.
BUILDING INDIA’S AI ENGINE
Developing Algorithms
for Economic Value
Modern AI expands what firms can do but its impact depends
on combining machine learning design with an
understanding of markets.
AI generated image
M
achine learning (ML) is the engineering
discipline of building systems that
learn from data and improve as they
are exposed to more of it. Creating ML systems
that deliver value requires two things: a) mathe-
matical formulation that is aligned with the true ob-
jective, b) sufficient data from which the algorithm
can learn patterns that steer its behavior toward
that objective.
Deploying modern AI systems is not very different.
Large pre-trained models come with prior human
knowledge and can perform multi-step tasks such
as ingesting documents, generating code and exe-
cuting workflows. Yet, leveraging these capabilities
within a firm depends on the same two require-
ments.
The advent of large datasets, advanced algorithms,
and modern AI has made a firm’s production func-
tion more malleable, widening the range of what it
can produce and whom it can serve. Consider a
bank providing working capital loans. It can now
draw on networks of payment transactions to as-
sess creditworthiness, enabling it to serve new cus-
tomer segments. This pattern extends across in-
dustries and functions. Creative applications of
data science and AI are expanding what firms can
do, with more such possibilities waiting to be built.
Importantly, reshaping a firm’s operating envelope
no longer requires building new factories or enter-
ing new markets; it can be done by analytics and
data science teams on an ongoing basis.
Despite this change, its potential remains largely
unrealized. Converting it into value depends on a
scarce skill: reading, at once, what the data and
systems can be made to yield and what that out-
put would be worth and to whom, then choosing
what to predict—the point at which these two
readings become a single decision. Take a data sci-
entist working at an insurance marketplace who is
asked to make a pipeline more effective. The firm
does not earn by predicting a single risk particu-
larly well; it earns by negotiating with insurers. As
a result, effectiveness is defined outside the data
basis which segment it can price better than its ri-
vals, which insurer is interested in that segment,
and which statistic carries weight when a contract
is struck.
Choosing what the pipeline computes is, therefore,
a modeling and a commercial decision. It is in this
single act that technical and business judgment be-
come inseparable, and only someone who holds
both perspectives can identify the move that a
purely technical or purely commercial specialist
would overlook.
Nikhil George
Visiting Assistant Professor
Plaksha University
08
1997
I Beat a World Chess
Champion
IBM's Deep Blue defeated world chess champion
Garry Kasparov in a six-game match. This was a
victory for computation and algorithm design: a
demonstration that machines could outperform
the finest human minds, although, in rule-
governed domains. But Deep Blue relied on brute-
force search and hand-crafted evaluation
functions only.
B U I L D
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure |
Research at Plaksha | Student Innovations
BUILDING INDIA’S AI ENGINE
Breakthroughs in the
Making
Research at Plaksha that uses AI to plan better cities and
build dependable robots.
Toolkit for Stubborn AI Problems
tion, University of Washington Seattle, Nagarro,
and the ATE Chandra Foundation ground the
research in stakeholder needs. The GeoAI lab
operates as an impact studio, moving from novel
methods to deployed prototypes, so that the next
generation of Indian cities grows smarter, greener
and more inclusive by design.
Anupam Sobti
Assistant Professor
Plaksha University
09
2012
Deep Learning Revolution
AlexNet, developed by Geoffrey Hinton's team,
outperformed every other system in the ImageNet
image recognition competition. This was the moment
deep neural networks stacked layers of learnable
transformations and demonstrated their power.
Computational power, large datasets and architectural
innovations came together to enable unimaginable
capabilities.
I could recognize faces, transcribe speech and translate
between languages. Deep generative models, systems
that could create new content began to emerge. But
deep learning required massive datasets and
extraordinary computational resources. Training large
models was expensive for most researchers.
C R E A T E
As cities expand and villages adapt to a changing
climate, planners are being asked to make faster
decisions on mobility, heat, water, housing and
livelihoods, often with incomplete data. Combin-
ing satellite imagery, street-level photographs,
GPS trajectories and citizen-generated reports
into machine learning models, the GeoAI lab is
turning raw geospatial signals into decision-ready
insight.
The lab organizes its work around four pillars. In
urban mobility, graph neural networks trained on
millions of GPS trajectories are reshaping bus
route design in Gurugram and will soon inform
Chandigarh's proposed 77 km metro in collabora-
tion with the University of Pennsylvania. In urban
climate, computer vision pipelines monitor cool
roof adoption across Indian cities, build auto-
mated tree inventories for Supreme Court
compliance and track water bodies using SAR
and optical imagery to improve flood resilience
with IIT Gandhinagar. In rural landscapes,
compact visual transformers enable crop moni-
toring and yield forecasting, while CaneAI
delivers farm-level alerts to sugar mills managing
lakhs of farmers. In AI4Gov, the team is co-
developing agentic systems with the Punjab
Urban Planning and Development Authority to
automate building plan approvals and with
Invest Punjab to streamline investor approvals
through retrieval-augmented AI.
Partnerships with Janaagraha, Raahgiri Founda-
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure |
Developing Algorithms for Economic Value | Student Innovations
A glimpse into the projects at Plaksha’s GeoAI Lab, from
bus-route design to flood control.
BUILDING INDIA’S AI ENGINE
Reliable Robots, Deployed in the Real World
Most people encounter AI through a chat interface.
The harder challenge is getting intelligence into a
body that moves through the physical world; it may
be less visible but is no less consequential. At
Plaksha's AI and Robotics Lab, that is exactly the
problem under investigation.
The lab's research is organized around a key
objective – reliability of robots. Three questions
drive it: how do you systematically teach robots
versatile motor skills? How do you get them to plan
reliably over a longer horizon in complex
environments? And how can they understand and
reason about the world around them? Versatile
motor skills, long-horizon planning, and embodied
reasoning define the lab's agenda.
Legged robots are the test case of choice. They are
harder to control but are far more versatile,
capable of climbing stairs, crossing rubble, and
adapting to surfaces. However, the dominant
training method, imitation learning, has a
weakness: models trained by copying expert
demonstrations do not generalize well. They are
brittle in novel scenarios and poorly grounded in
physics. The lab's research targets these gaps.
On the locomotion side, we have developed and
deployed a quadruped locomotion policy trained
using curriculum-based reinforcement learning in
Isaac-Sim. The work tests multiple architectures,
blind policies relying only on proprioception,
terrain-aware policies with explicit geometry
sensing and adaptive variants that adjust to mid-
episode friction changes, establishing clear
performance tradeoffs between them. Physical
deployment on a quadruped robot has been
achieved, with a structured sit-to-stand transition
added to the pipeline. Rahath Malladi, a fourth-
year BTech student, extends this work toward
multi-embodiment generalization: a single policy
trained to walk Anymal, Spot, and the H1
humanoid across a shared simulation
environment.
The lab also works on the planning layer. Bharat
Jain, a third-year BTech student, has built a VLM-
based autonomous navigation system that
connects Claude to ROS via MCP, enabling a robot
to explore an indoor environment, build occupancy
map, label it using a vision-language model, and
navigate to any object a user names in natural
language. This framework has been shown to work
in the Gazebo simulation environment. Separately,
Arunachalam Vijayanand, a second-year BTech
student, is fine-tuning LLMs for direct robot control
using reinforcement learning techniques. The goal
here is to develop a framework to tune LLM or VLM
as that would enable them to be grounded better
in the physics that underpins the robot and its
interactions with the real world. In some initial
experiments, the lab was to train an LLM to reach a
52% goal-completion rate on a drone maze task,
matching state-of-the-art performance.
Cutting across all three streams is a model-based
learning approach being developed in collaboration
with UC San Diego, and UC Irvine. World models by
their very nature are built to replicate the physics
of the real world and they in principle can be
learned by an agent itself by self-exploration. Such
a world-model, in principle, removes the grounding
problem and affords the agent the ability to reason
and plan by simulating the world model. The lab
has shown that a world model learned entirely
through self-exploration without any prior data can
support policy learning that is significantly more
efficient compared to existing state-of-the-art
model-based learning approaches and holds up
under out-of-distribution perturbations.
The work is incremental in the best sense: each
project addresses a specific, well-defined gap.
Together, they are laying groundwork for robots
that can be deployed in the real world.
Deepan Muthirayan
Assistant Professor
Plaksha University
10
2014
I Learned to Imagine
Two landmark deep generative models arrived.
Generative Adversarial Networks (GANs), introduced by
Ian Goodfellow and colleagues, pitted a generator
against a discriminator in an adversarial game, each
improving the other until the generator could produce
images indistinguishable from reality. Variational
Autoencoders (VAEs) offered a complementary
approach: learning compact latent representations of
data and sampling from them to generate diverse new
content.
However, GANs were difficult to train and prone to
mode collapse. VAEs produced diverse results, often
with unwanted blurriness.
C R E A T E
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure |
Developing Algorithms for Economic Value | Student Innovations
BTech students at work in the ARK Robotics Lab
BUILDING INDIA’S AI ENGINE
C R E A T E
AI in Practice: Student
Actions that Matter
Plakshans are applying AI to drive research
breakthroughs, develop impactful technologies and solve
complex societal challenges.
Supporting 22 Indian languages along with English
and spanning speech, translation, chat and agentic
workflows, the flagship 30B and 105B models of
Sarvam AI form the backbone of widely used
products like Samvaad and Indus. But beyond the
technology, they make a statement: India should
not perpetually be a consumer of AI infrastructure
built elsewhere, optimized for someone else's
languages, institutions and context.
When I graduated from Plaksha in 2025, I came to
understand that sovereign AI is not just a patriotic
talking point; it is an engineering and research
discipline. Not just understanding Indian linguistic
nuance, but working reliably across diverse
domains and competing with open-source giants
at the state-of-the-art. We are the first to do this at
this scale, in a severe time crunch, solving
problems no one has solved here before. That
difficulty is exactly what makes it worth doing!
So, how does a model get built? First, gather
everything the internet has ever seen. Clean it
because bad data bites. Toss it all onto a mountain
of GPUs. Watch for spikes. Run evaluations along
the way to make sure it is learning. Teach it
through SFT and RL. Finally, sprinkle in some
personality. Do all of that and you have made a
model.
However, building a model is a long sequence of
small, unglamorous decisions. You start with a
clean hypothesis. Then reality adds edge cases.
Then the edge cases add more edge cases.
Eventually it stops feeling like ‘training a model’
and starts feeling like trying to keep a very large,
expensive, intelligent system honest. Being one of
the younger engineers in the room, with the least
context, turned out to be a surprisingly efficient
learning algorithm for me. Vague thinking gets
exposed fast. Patience stops being a virtue and
becomes a job requirement.
The models are live, in real products, serving real
users across India. But what I find more
compelling than where we are is how we got here
and how we keep going. The team is continuously
identifying failure modes, running evaluations
and iterating in response to what those
evaluations reveal. It is a compounding loop: find
where the model is wrong, understand why, fix it
carefully and make sure the fix did not break
something else.
India's AI future will not arrive in one dramatic
moment; it will be built exactly like this, iteratively
by people willing to do the work.
Soham Petkar
Machine Learning Engineer
Sarvam
BTech Class of 2025
11
2017
Transformers Change
Everything
Researchers at Google published ‘Attention Is All You
Need,’ introducing the transformer architecture.
Transformers could attend to all parts of a sequence
simultaneously, capturing long-range dependencies
with precision. This architecture became the
foundation of every large language model that
followed: vision transformers, multimodal models and
generative systems beyond language.
I could now hold context across long documents and
translate between languages. At scale, it required
enormous computational resources and raised
questions about energy consumption, bias embedded
in training data and the ethical governance of systems.
C R E A T E
India's AI, Built at Sarvam
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure |
Developing Algorithms for Economic Value | Research at Plaksha
(Bottom right, seated) Soham Petkar, BTech 2025,
with Sarvam AI’s team of young engineers.
BUILDING INDIA’S AI ENGINE
Utilizing AI Across Healthcare, Education and
the Environment
At Plaksha, I discovered that the boundary
between a classroom and a research lab is
intentionally thin. Courses are not just academic;
they are launchpads. Through Geospatial Data
Science and ILGC, I used AI for remote sensing to
explore satellite-based lake monitoring and urban
flood detection using radar imagery. As part of the
Human-Technology Interaction course, I built
AwareLLM, where AI assistants adapt to a user's
cognitive and physiological state and tailor
proactive responses to boost workplace
productivity. What I eventually discovered was that
AI is a powerful force multiplier for solving every
domain problem.
This exposure gave me the opportunity to explore
my passion for healthcare through the lens of AI.
In Plaksha's Deep Learning course, my team and I
worked on lesion segmentation in neonatal
hypoxic-ischemic encephalopathy (HIE), a severe
neurological condition affecting roughly 1 in 5,000
newborns. The challenge of identifying and
localizing these lesions is immense: HIE lesions are
often very small, scattered unpredictably and lack
global normative references. To address this, my
team and I designed a computationally efficient AI
framework that introduced spatial priors encoding
where abnormalities are likely to appear, paired
with a boundary- and region-aware loss function
to prevent training collapse under extreme class
imbalance. What began as a course project soon
evolved into a research collaboration with faculty
at Harvard Medical School. The project was also
accepted to MICCAI 2025, a premier conference at
the intersection of medical imaging and AI. We
presented our work to an international community
in Daejeon, South Korea, where we were among
just a handful of undergraduates alongside senior
researchers, clinicians and industry experts.
Similarly, in the Machine Learning and Pattern
Recognition course, I tackled pneumothorax
segmentation in chest X-rays. By using deep
learning-based bone suppression to remove
obscuring ribs and clavicles from the images, I
learned that AI model performance in pixel-level
disease localization improved significantly and
consistently. This work, also developed in
collaboration with researchers at Harvard Medical
School, earned an oral presentation at IEEE ISBI
2026 in London.
Beyond the lab, my internships brought exposure
to a deployable scale. At Wadhwani AI, I worked on
SEAP, an AI-powered spoken English assessment
and practice platform supported by the Meta
Llama Impact Grant and built for early-career
professionals, first-generation English learners and
rural students across India. The platform analyzes
a learner's pronunciation at the phoneme level
and adaptively generates personalized reading
content targeting specific areas of difficulty,
making English learning accessible at scale.
Integrated into SWAYAM, India's national digital
learning platform, the system has powered over
60,000 assessments for more than 27,000 learners
in its initial phases. A fully funded research
internship at MBZUAI in Abu Dhabi has given me
the opportunity to explore world models for
medical vision-language understanding, with the
goal of building models that learn deeper
semantic representations of clinical concepts.
What ties all these experiences together is the
belief that AI, when built thoughtfully, can operate
at the intersection of science and human need.
There is something humbling about realizing that
work which began as a classroom project can find
its way into global research conversations and into
the hands of thousands of learners.
Amog Rao
Research Fellow
Microsoft Research
BTech Class of 2026
12
2022
The Generative AI Era
ChatGPT became one of the fastest-adopted
technologies in history. Models like Gemini, LLaMA, and
DeepSeek extended capabilities across reasoning, code
and multimodal understanding. Prompt engineering
emerged as a discipline.
I moved from research labs into everyday life, assisting
with writing, programming, scientific discovery,
education and creative work. But my rapid adoption
surfaced challenges: hallucination, reasoning errors,
susceptibility to adversarial prompts and black-box
opacity. Questions of misinformation, job
displacement, bias, intellectual property and
governance became urgent.
C R E A T E
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure |
Developing Algorithms for Economic Value | Research at Plaksha
BUILDING INDIA’S AI ENGINE
AI for Scientific Discovery
There is a particular quality of attention that settles
over a room when a leading scientist describes not
what they know, but what they cannot yet explain. I
experienced this during ConfAI. Surrounded by
researchers and industry leaders, I was captivated by
the vastness of what remained unsolved. I realized
that the most exciting questions often existed
between disciplines.
At Plaksha, I could study AI, biology and healthcare in
an environment that encouraged me to pursue
questions across domains. Each of my research
experiences was an exploration of that idea: how can
AI help us uncover structure in complex biological
systems that remain partially understood?
I first encountered this question during my research
internship at Brown University through my work on
AlphaFold2 and protein conformational ensembles.
AlphaFold transformed structural biology by
predicting protein structures with accuracy, yet
proteins themselves are fundamentally dynamic.
Their function often emerges from transitions
between many possible states.
Through the development of FastConformation, I
investigated whether AlphaFold's internal
representations captured this underlying
conformational landscape or whether they collapsed
toward structures already seen during training. The
work revealed evidence of abstraction and
memorization, raising a deeper question: when does
a model genuinely learn the governing principles of a
biological system, and when is it recalling examples
from experience? For me, this became an early
lesson that scientific discovery requires models
capable of representing possibilities rather than
producing predictions.
That perspective led me to the Image, Informatics &
Intelligence (i3) Research Lab at Harvard Medical
School, where I worked on pediatric brain foundation
models. The goal here was to learn representations
of neurodevelopment. The pediatric brain is a
continuously evolving system, shaped by age,
development and immense biological variability.
Using multimodal self-supervised learning, I sought
to build models capable of capturing normative
developmental trajectories while adapting to rare
neurological disorders with limited supervision. A
particularly challenging aspect involved missing
imaging modalities, a common reality in clinical
practice. Building representations that remained
faithful despite incomplete observations reinforced a
broader scientific principle: robust models should
preserve the underlying structure of a phenomenon
even when the data is imperfect.
Most recently, at Microsoft Research India, I explored
how foundation models can bridge perception,
language and clinical reasoning. Radiology is
inherently multimodal. Therefore, a radiologist
needs to connect anatomical location, morphology,
imaging appearance and clinical language into a
coherent explanation. My work focused on phrase-
grounded segmentation, clinical concept
decomposition and grounded radiology report
generation, with the aim of developing systems that
can align free-text clinical descriptions with precise
locations in three-dimensional CT and MRI scans. By
decomposing continuous representations into
clinically meaningful concepts, capturing
appearance, morphology, scale and anatomical
context, I moved closer to models that can explain
how those findings relate to the underlying evidence.
Looking ahead, I am excited to explore the next
frontier of AI through agentic memory and visual
perception as part of the Microsoft 365 Copilot
research team.
Ananya Shukla
Research Fellow
M365 Research, Microsoft
BTech Class of 2026
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure |
Developing Algorithms for Economic Value | Research at Plaksha
13
2025
I Learned to Reason
A new generation of models, beginning with
OpenAI's o1 and followed by DeepSeek's R1
paused and worked through problems step by
step.
DeepSeek's R1 demonstrated that frontier
reasoning could be achieved at a fraction of the
cost and compute of American models. It was
built by a small team in Hangzhou and open-
sourced to the world.
C R E A T E
BUILDING INDIA’S AI ENGINE
Held at the Plaksha campus on Mar 11, 2026, the event
brought together eminent voices from academia and
industry. It featured Vijay Kumar, Nemirovsky Family Dean
of Penn Engineering, UPenn; Chris Callison-Burch, Professor
of Computer and Information Science, UPenn; Ravi
Radhakrishnan, Herman P. Schwan Department Chair of
Bioengineering, UPenn; and Neeraj Aggarwal, Vice Chair,
Asia Pacific, BCG. The conversation explored how AI is
reshaping industries and scientific research.
*As of Aug 2026
This edition of 1729 presents a new take on mandala art, an ancient
tradition known for intricate geometric patterns and for representing
balance, harmony and interconnectedness. Traditionally, mandalas
are crafted through detailed motifs by hand; they embody
mindfulness, creativity and the pursuit of knowledge. The new
design draws on India’s rich artistic heritage and brings technology
into the process. Digital design tools and algorithms were used to
build the geometry, so that each pattern could be generated, refined
and repeated with precision. The result reflects the vision at Plaksha
of fostering innovation.
Plaksha University
Alpha, Sector 101, IT City Road, Sahibzada Ajit Singh Nagar, Punjab 140306
42
Full-time
Faculty
12
Visiting
Faculty
800+
Student
Body
227
Executives &
Teaching Fellows
51
Research
Staff
420+
Alumni
Join hands with Plaksha: We are keen to collaborate
with faculty members from within and outside Plaksha
to push the boundaries of research and innovation.
Virtual tour: Explore the vibrant campus of Plaksha, its
state-of-the-art facilities and collaborative learning
spaces, from the comfort of your home.
14
The Harder Conversation
The European Union's AI Act, the first ‘comprehensive
legal framework for AI’, began its phased
implementation through 2025, banning the highest-risk
applications and imposing transparency obligations on
the models reaching the largest populations. 60
countries endorsed a declaration calling for inclusive
and ethical AI development. The question of who gets to
set the rules and whose values those rules encode
remains bitterly unresolved.
And then comes the bigger question: whether my
benefits justify the costs. In 2024, global data centers
consumed roughly 1.5% of all electricity used on Earth.
By 2030, that figure will more than double. So, every
query I answer, every image I generate, every reasoning
chain I trace carries an energy cost that my users rarely
see.
Hasrat Sandhu and
Rashmi Ranial
CTLC Team