CV

Research Assistant (ML Security and LLM Privacy)

Dartmouth College

Advisor: Prof. Shawn Shan

• Co-developed CHASE (Correlated Hierarchical Adaptive Search for Extraction), a Gaussian-process bandit framework for query-efficient memorization extraction from instruction-tuned LLMs.
• Found that the prompt space for memorization extraction has a low-rank, kernel-correlated structure, enabling effective prompt optimization with as few as 100 API queries per target passage.
• Evaluated across 9 instruction-tuned LLMs from 7 model families, including production APIs, with up to 2.9x higher extraction success than random search and over 10x gains under seven defense strategies.
• Showed that ordinary academic and continuation prompts consistently outperform adversarial jailbreaks for memorization extraction, exposing limits of prompt-level safety defenses.

Founder and Lead Researcher

Adhāra AI Labs

Independent AI research lab

• Founded an independent research lab focused on generative AI, LLMs, retrieval-augmented generation, and deep learning systems.
• Research outputs include PyraFuseNet (ICIAI), SLMs as Compiler Experts (NeurIPS 2025 ML for Systems Workshop), RegimeNAS, MorphNAS, and a low-latency jailbreak prevention framework.

AI Research Intern

Lossfunk

Low-resource language modeling

• Built an Indic LLM pipeline for extremely low-resource languages (Tulu, about 0.001% of typical training data volume).
• Developed a hard-negative-constraint framework that reduced catastrophic language leakage from 80% to 5%, showing that explicit prohibitions outperform positive instructions.
• Resulting paper accepted at the EACL 2026 Workshop on Language Models for Low-Resource Languages (LoResLM).

Intent-Based Network Management System (3GPP SON)

Nokia

Industry internship

• Designed an intent-based self-organizing network system that translates operators’ natural-language goals into 3GPP-compliant intents.
• Built an LLM-powered, retrieval-augmented intent classification pipeline mapping natural-language inputs to structured intent schemas.
• Architected an orchestration layer with synchronous and asynchronous execution, conflict detection, duplicate suppression, and intent expiry management.

Contributor

Google Summer of Code 2025, University of California Santa Cruz

Billion-scale ANN benchmarking infrastructure

• Built billion-scale vector embedding benchmarks (768, 1024, and 2048 dimensions) from open-source codebases for realistic evaluation of approximate nearest neighbor search.
• Developed infrastructure for evaluating ANN algorithms at scale.

Undergraduate Research Intern (Deep Learning)

Indian Institute of Technology, Indore

Advisor: Prof. Nagendra Kumar

• Designed the KASPER framework for PDF malware detection, reaching 99.5% accuracy and maintaining it under FGSM and PGD adversarial attacks.
• Built a malware injection pipeline for realistic payload embedding, and led benchmarking and explainability analysis using Kolmogorov-Arnold Networks.
• Work formed the core technical implementation of a paper in Applied Soft Computing.

Teaching Assistant, Machine Learning and Deep Learning

PES University

Teaching

• Taught and mentored a cohort of 75+ students; prepared lecture slides, delivered tutorials, and ran weekly hands-on labs.
• Designed and graded assignments, lab exercises, and hackathons.

Subject Matter Expert, Deep Learning

PESU I/O

Teaching

• Co-instructed a 30-hour Deep Learning from Scratch course for 40 students, delivering 20 hours of in-person instruction on neural networks, CNNs, RNNs, LSTMs, GANs, and explainable AI.

Community Mentor and Technical Speaker

Independent

Service

• Mentored 50+ teams in hackathons, research projects, and WiDS Datathon competitions.
• Delivered technical talks on DSPy and LoRA fine-tuning at FOSS United and GDSC events.

Head of Technology

Entrepreneurship Club, PES University

Leadership

• Led a 10+ member technical team; built the club website and ran technical operations for a 100+ member organization.

Webmaster

IEEE Robotics and Automation Society, PES University

Service

• Managed technical infrastructure and web presence for the university robotics organization.

Core Member

Google Developer Student Clubs and Hacker Space, PES University

Service

• Organized technical workshops and community events promoting free and open-source software.

PhD

Dartmouth College

Computer Science

• Advisor: Prof. Shawn Shan, JASPR Lab
• Research interests: trustworthy machine learning, AI security and privacy, large language models, memorization and generalization theory, adversarial machine learning, responsible AI systems

BTech

PES University

Computer Science Engineering

• Relevant coursework: Machine Learning, Deep Learning, Generative AI, Linear Algebra and its Applications, Statistics for Data Science, Engineering Mathematics I & II, Data Structures and Algorithms, Operating Systems, Computer Networks, Distributed Systems

Deletion Scars: Inferring Deleted Artist Styles from Paired Diffusion Models

ECCV 2026 Workshop on Unlearning and Model Editing (U&ME)

P. Devadiga

Resistant to Lawyers, Defeated by Disagreement: Evaluation Blindspots in Legal Language Models

ICML 2026 Workshop on AI for Law (AI4Law)

P. Devadiga, A. Lakshman

Making Large Language Models Speak Tulu: Structured Prompting for an Extremely Low-Resource Language

EACL 2026 Workshop on Language Models for Low-Resource Languages (LoResLM)

P. Devadiga, et al.

GUARDIAN: Multi-Agent Defense System for Large Language Model Security

International Conference on Artificial Intelligence and Information (ICIAI), Waseda University, Japan

P. Devadiga, et al.

SLMs as Compiler Experts: Auto-Parallelization for Heterogeneous Systems

NeurIPS 2025 Workshop on Machine Learning for Systems

P. Devadiga, et al.

PyraFuseNet: Dual-Path Network for Resource-Constrained Vision

International Conference on Artificial Intelligence and Information (ICIAI), Nanyang Technological University, Singapore

P. Devadiga, et al.

SAMVAD: Multi-Agent Framework for Modeling Judicial Reasoning in Indian Jurisprudence

arXiv preprint arXiv:2509.03793

P. Devadiga, et al.

RegimeNAS: Regime-Aware Neural Architecture Search for Financial Trading

arXiv preprint arXiv:2508.11338

P. Devadiga, et al.

MorphNAS: Differentiable Neural Architecture Search for Multilingual Named Entity Recognition

arXiv preprint arXiv:2508.15836

P. Devadiga, et al.

Amazon AI-ML Scholar

Amazon Web Services

Selected among the top 1,000 AI/ML students in India.

Oxford Machine Learning Summer School

University of Oxford

Selected participant.

Cohere AI Summer School

Cohere

Selected participant in a program on large language models and NLP.

Winner, Cisco ThingQbator Hackathon 6.0

Cisco

National hackathon win among 1,000+ competing teams.

Merit Scholarships (6x)

PES University

MRD Scholarship (2x), CN Rao Scholarship (2x), and Distinction Scholarship (2x) for top 5-20% academic performance.