I make AI efficient enough to run on small, low-power devices, so it keeps working where the cloud
can't reach.
NowResearch and Development Intern (Erasmus+) University of Ljubljana, Slovenia
From 2027Visiting Researcher Cognitive Interface Lab, Nanyang Technological University Singapore
StudyingBachelor of Technology in Computer Science and Engineering Shri Mata Vaishno Devi University
About
I'm currently a Research and Development Intern under the Erasmus+ programme at the
University of Ljubljana, Slovenia, working with Prof. Veljko Pejović on carbon-aware
scheduling for large language model inference. In January 2027, I will join the Cognitive
Interface Lab at Nanyang Technological University Singapore as a Visiting
Researcher.
My research focuses on deploying foundation models at the edge of autonomous and embodied systems.
Using hardware-software co-design, model compression, and
edge-native sensor fusion, I work on making models run in real time on
microcontrollers and single-board computers, without depending on the cloud. I also care about
keeping these systems robust, secure, and interpretable under real-world constraints.
I am pursuing a Bachelor of Technology in Computer Science and Engineering at Shri Mata Vaishno Devi
University in Jammu and Kashmir, India.
Research interests
i.Hardware-software co-design
ii.Embodied AI
iii.Human-computer interaction
iv.Model compression for edge AI
v.Sustainable and carbon-aware computing
Education
Shri Mata Vaishno Devi University
Bachelor of Technology in Computer Science and Engineering
2025 – 2029
Nanyang Technological University Singapore
Visiting Research Student (non-graduating), College of Computing and Data Science
2027 – 2028
News
Nominated Student Coordinator for the Office of Alumni & International Affairs at Shri Mata Vaishno Devi University.
Selected as a 2026 Ambassador for Thinking About Thinking, a nonprofit community on AI, neuroscience, and the foundations of intelligence.
Selected as a Global Youth Ambassador for Theirworld, advocating for education in displaced and conflict-affected communities.
Launched pjk.co.in as Technical Director of PAT & J.K. Co.
Started aryaveer.mind, an explainer series on AI research without the jargon.
Began an Erasmus+ traineeship at the University of Ljubljana, Slovenia, with Prof. Veljko Pejović.
Invited to the Cognitive Interface Lab at Nanyang Technological University Singapore for a research attachment from January 2027.
Research Experience
Visiting Researcher Incoming
Singapore
Cognitive Interface Lab, Nanyang Technological University SingaporeJan 2027 – Jan 2028
A 12-month attachment with the College of Computing and Data Science on edge-native cognitive AI:
closed-loop wearable systems that turn biosignals into low-latency feedback for human-computer
interaction.
The central problem is compression: taking architectures built for accelerator-scale compute and
finding what can be cut, quantised, or restructured to fit strict on-chip memory and power budgets
without losing the signal that matters.
Research & Development Intern (Erasmus+) Ongoing
Ljubljana, Slovenia (remote)
Faculty of Computer and Information Science, University of Ljubljana, SloveniaSep 2026 – Dec 2026
With Prof. Veljko Pejović on resource-aware mobile computing and sustainable edge AI.
Building a carbon-aware scheduler that places large language model queries across a device, edge,
and cloud continuum. See the project below.
Research Intern
Jammu and Kashmir, India
Shri Mata Vaishno Devi UniversityMay 2026 – Jul 2026
Behaviour-based detection of malicious USB charging sessions. See the project below.
Projects
Sustainable computing Ongoing
Carbon-Aware Scheduling for Large Language Model Inference
University of Ljubljana, Slovenia, with Prof. Veljko Pejović · 2026 – present
Hardware tiers
Device
Apple M4 laptop
measured
Edge
RTX 3070 desktop
next
Data centre
NVIDIA A100
next
Figure 1. Where a query can run. Click to enlarge.
Interactive queries can't be postponed until the grid is cleaner. So instead of shifting
when a query runs, this scheduler shifts where it runs (on the user's device, an
edge server, or a data centre) and how (model size and quantisation). It reduces emissions
while keeping time-to-first-token, time per output token, and answer quality within set limits.
The work rests on a measurement harness that profiles Qwen2.5 models served with llama.cpp on
each hardware tier.
Early findings from the device tier (Apple M4 laptop)
≈10×Generating a token costs about ten times more energy than reading one.
−53%Batching 16 requests cuts energy per query, but waiting time to the first token rises twelvefold.
≈4×From Slovenia, choosing where a query runs within a 30 ms latency budget saves about four times more carbon than delaying it by 24 hours.
0.3%Energy per query reproduces across three days to within 0.3% variation.
Next: edge (RTX 3070) and data-centre (A100) measurements, the cost model, and the scheduler itself.
A detection pipeline and telemetry simulator for juice jacking, keystroke injection,
class-swapping devices, and concealed hardware, built on 60,000 simulated sessions. On chargers and
devices it never saw, the detector reaches 80% recall at a 1% false-positive budget, with high
variance across seeds. It scores 98% on familiar hardware, against 77% for a rule baseline. The
scorer exports to 27 KB of dependency-free C. All results are on synthetic data, and the project
audits and retracts an earlier version's inflated numbers.
Gradient boosting · SHAP · ONNX · C export
Edge AI
Nida-e-Watan: Offline Edge Inference Cluster
2025 – present
An offline-first AI system across 12 Raspberry Pi 4 nodes, serving villages with about two hours of
connectivity a day. Quantising and pruning multilingual BERT and IndicBERT shrank the model from
1.1 GB to 260 MB while keeping 91% task accuracy. On-device classification takes 380 ms, and
conflict-free replicated data types keep 99.2% data integrity between uplinks.
Raspberry Pi · ONNX Runtime · quantisation · pruning
Edge AI · Wearables
Eco-Smart Wearable Biosensing Patch
2026
A cloud-free biosensing patch built on an Arduino Nano with electrochemical electrodes and
temperature sensors. A time-series classifier runs on the microcontroller itself, within its memory
limits, to flag physiological anomalies in real time.
A ResNet-18 using five spectral bands beat one using only colour (97.76% vs 97.06%), checked across
random seeds and explained with Grad-CAM. The classified terrain then fed an A* router, where convoy,
recon, and stealth agents re-plan around soil moisture, visibility, and a simulated rainfall
event.
A 40-page monograph on internal displacement, economic stagnation, and technological exclusion in the
Pir Panjal and Chenab Valley regions of Jammu & Kashmir. It examines how the 1990s migrations still
shape the lives of the region's displaced youth, and what technology and AI governance could do.
aryaveer.mind
2026 – present
An AI explainer series on Instagram. Each post is a fact-checked, jargon-free carousel on a piece of
AI research or news, from Attention Is All You Need to Google's plan to put AI chips in
orbit.
Essays
Medium · LinkedIn
10+ essays on AI policy, digital inclusion, natural language processing for low-resource languages,
and edge computing for social good. A recent one asks whether model compression costs low-resource
languages more than English.
Service & Leadership
Technical Director & Head of Quality Assurance2026 – present
PAT & J.K. Co.: built the group's web platform, starting with Phulhaar Gallery
Student Coordinator2026 – present
Office of Alumni & International Affairs, Shri Mata Vaishno Devi University
Ambassador2026 – present
Thinking About Thinking
Global Youth Ambassador2026 – present
Theirworld
Citizen Scientist2026 – present
AI/ML Analysis Working Group, Open Science Data Repository
Founder2025 – present
Project Nida-e-Watan: digital literacy and welfare access for displaced families in the Chenab Valley, in Hindi, Urdu, and Dogri
Honours
TryHackMe, top 1% globally(top 500 worldwide, top 50 in India)
Machine Learning Specialization · Stanford Online
Deep Learning Specialization · DeepLearning.AI
Mathematics for Machine Learning · DeepLearning.AI
AI, Justice, and the Rule of Law · University of Oxford
Technical Skills
Programming
Python, C++, JavaScript, SQL, Bash, Git, Linux
Machine learning
PyTorch, TensorFlow, Hugging Face Transformers, scikit-learn
Deep learning
Transformers, BERT variants, CNNs, LSTMs, ResNet
Model optimisation
4-bit and 8-bit quantisation, pruning, knowledge distillation, LLM fine-tuning
I'm always glad to talk about edge AI, compression, carbon-aware computing, or
low-resource language technology. Write to me at
aryavsparihar@gmail.com.
