• Attendance of RISC 2025 Participants
• Distribution of Participant Kits

HoD addresses participants and guests about RISC 2025 event and gives overview of Computer Science and Engineering in IIT Bombay

Planar Graphs in Real Life - The Problem of Land Records

Abstract: We look at the problem of Land Records for farm lands - maintaining an accurate match between actual land use and its record of ownership. For the first, there is now an option to time-consuming land surveys - viz. segmentation of satellite or drone images. For the second, we must look at how the government maintains land records, which is as a number of land parcels or polygons in imaginary space in GIS. The question of matching the two then becomes polygon fitting to segmented data. We look at theoretical, algorithmic and practical problems which arise in this process and describe our solutions. This work is part of a project with Govt. of Maharashtra and uses past independent collaboration with Google Research.

Speaker: Prof. Milind Sohoni

Beyond In-Context Learning: Rethinking Few-Shot Adaptation for Structured Tasks

Abstract: Many applications require few-shot adaptation for structured prediction tasks, such as semantic parsing for custom APIs, translation involving low-resource languages, and Text-to-SQL for private databases. These can be framed as structured sequence-to-sequence problems. While In-Context Learning (ICL) is the dominant approach for adapting large language models (LLMs) to new tasks, we show that it has fundamental limitations in this setting. Through a mechanistic analysis with synthetic formal languages, we attribute the failure to difficulty of in-context learning input-output alignments in structured tasks. For light-weight adaptation on such tasks, we will discuss alternative strategies, including case-based reasoning on trees, in-context fine-tuning, paired decomposition of input-output sequences, and grammar-constrained decoding.

Speaker: Prof. Sunita Sarawagi

Bio: Sunita Sarawagi researches in the fields of databases, machine learning, and applied NLP. She is Institute Chair Professor in the Computer Science Department and was the founding head of the Center for AI at IIT Bombay. She got her PhD in databases from the University of California at Berkeley and a bachelor’s degree from IIT Kharagpur. She has also worked at Google Research, CMU, and IBM Almaden Research Center. She is an ACM fellow, was awarded the Infosys Prize in 2019 for Engineering and Computer Science, and the distinguished Alumnus award from IIT Kharagpur. She has several publications including notable paper awards at ACM SIGMOD, ICDM, and NeuRIPS conferences.

Room No: CC-101

Speaker	Talk Title
Vedant Kalbande	Intelligent Caching for Datacenter Applications
Samipa Samanta	Dynahints: Silent Threshold Signature for Dynamic Setting
Narjis Asad	BIStereo: Using NLI to Detect Skin Complexion Biases in LMs
Abhik Bose	P4 programmable data plane for complex applications
Omkar Vijaykumar Tuppe	Verification of Transactional Programs under Serializability

Room No: CC-103

Speaker	Talk Title
Aditya Anand	Program Analysis for Managed Runtimes in Presence of Dynamic Features
Gargi Bakshi	Limiting Disease Spreading in Human Networks
Lokesh N	Robust Root Cause Diagnosis using In-Distribution Interventions
Santhosh Kumar M	Flyt: Software-defined elastic GPU endpoints
Devdan Dey	Online ε-Net for Geometric Objects

Room No: CC-105

Speaker	Talk Title
Prathamesh Navale	Adaptive Software Prefetching for Recommendation Systems at-scale
Himanshi Singh	A Convex Hull approach to a sub-quadratic algorithm for Motzkin Rabin Theorem
Sandarbh Yadav	Portfolio Optimisation using Policy Search
Heinsamding Thou	Will you choose reliability over low-delay in VoIP calls over poor networks?
Soham Joshi	Monotone Submodular Multiway Partition

Participants will have lunch and engaging discussions with CSE Department folks

Formal Methods: Theory and Practice

Abstract: I will present my research journey in formal methods, with focus on the theory of string transformations. The discussion on string transformation will touch upon diverse topics like logic, automata, algebra, constraints. Finally I will discuss the importance of formal methods in industry, the kind of problems we face in the hardware industry and the need for continuous innovation in this area.

Speaker: Dr. Vrunda Dave

Bio: Vrunda Dave received her PhD from CSE department, IIT Bombay under the guidance of Prof. Krishna S. For her thesis on “Some Fundamental Problems and Applications of Word Transformations”, she received Honorable Mention for ACM India Doctoral Dissertation Award’22, and an Excellence for Ph.D Research award from CSE Department, IIT Bombay. She was also a finalist for the E. W. Beth Outstanding Dissertation Prize’22. She works as an Architectural Formal Verification Engineer at Intel.

A Journey to the Post-Quantum-Cryptography

Abstract: Quantum computers are here. While there are ongoing debates on to what extent it is going to replace our good old classical computers, there is no doubt about the fact that our security will surely go for a toss. This is due to Shor’s algorithm and its successors, which can solve the prime factorization and discrete log problems efficiently — two problems that define modern cryptography. Given this fact, the world is rapidly moving towards “Post-Quantum-Cryptography” (PQC), which can withstand quantum attacks. This talk will give a (extremely informal) introduction to Lattice-Based Cryptography, which is going to run on all of your devices in the next few years.

Speaker: Prof. Sayandeep Saha

Panel Topic: Shaping Tomorrow’s Workforce

Panelists:

1. Mr. Aman Goel
2. Dr. Nagarajan Natarajan
3. Dr. Priyanka Naik
4. Prof. Shivaram Kalyanakrishnan
5. Dr. Vineet Nair
6. Dr. Vrunda Dave

Panel Moderator: Prof. Varsha Apte (HoD, Department of Computer Science and Engineering)

CSE Students will present their ongoing research work through Posters, with refreshing tea and snacks for everyone

Robust Customization of Large Language Models

Abstract: Pretrained large language models are increasingly achieving impressive zero-shot accuracies on challenging benchmarks. At the same time, customizing these models to perform well in specific settings is still crucial, e.g. fine-tuning models on private codebases or aligning models to application-specific requirements. In this talk, I’ll present an overview of the area of customizing LLMs that spans prompt engineering, fine-tuning, as well as post-training alignment. In particular, I’ll cover some of our recent ML work on training and aligning models to promote out-of-distribution generalization and robustness to noise in the training data.

Speaker: Dr. Nagarajan Natarajan

Bio: I am a Principal Researcher at Microsoft Research India. Over the last several years at MSR India, I’ve worked on a broad slate of machine learning problems at the intersection of AI and software engineering, AI and systems, as well as in learning theory and optimization. My current interests are largely shaped by the big challenges we face today as we increasingly deploy black-box Large Language Models in real systems like Co-pilots. I’ve collaborated with several researchers, scholars, and students who have shaped my thinking over the years. I owe a lot in particular to Prateek Jain, who mentored me at MSR India in the initial few years, to Prof. Ambuj Tewari who was instrumental in my formative years of research career, and to Prof. Inderjit Dhillon, my PhD advisor at the University of Texas at Austin.

Demystifying System Research

Abstract: System research is the core for building any great product or technology asset. We will explore how to get started on your research journey in systems and career opportunities in this space.

Speaker: Dr. Priyanka Naik

Bio: Priyanka is a Research Scientist at IBM India Research Lab and is part of the GenAI platform team. Priyanka’s prior research explored the space of cloud deployment, networking and observability of microservices such as network functions and extending the same for multi cloud deployments. Prior to joining IBM, she received her Ph.D. from IIT Bombay, India. She is a co-author to a cloud networking book and has also recently been appointed as an ACM Eminent Speaker and looks forward to connecting with institutes to collaborate on tutorials and talks.

Room No: CC-101

Speaker	Talk Title
Mandru Suma Sri	Machine Learning based Attack on Learning with Errors Problem
Prateek Garg	From Search to Sampling: Generative Models for Robust Algorithmic Recourse
Arhaan Ahmad	Sensitivity Analysis of Decision Trees
Anubhav Bhatla	OASIS: Enabling Provably Secure Randomized Caches at Ultra-Low Cost
Darshan Prabhu	Spicing Up ASR with LLMs for Low-Resource Languages

Room No: CC-103

Speaker	Talk Title
Supriya Bhide	Static Single Assignment for Compilers
Karthika N J	Leveraging Syntactic and Semantic Unity of Indian Languages for NLP
Rahul Kapur	Automated Functional Synthesis: Towards Correct-by-Construction Systems
Prerna Priyadarshini	The Other Side of Dynamic Cache Partitioning
Prerak Contractor	From Theory to Practice: Developing efficient tools

Room No: CC-105

Speaker	Talk Title
Saurav Chaudhary	TABuddy: AI-Assisted Grading Tool for Introductory Programming Assignments
Krishna N Agaram	Quantum State Preparation with Reinforcement Learning
Namrita Varshney	ViKriTi: Glitches in Tree Ensemble Models
Abhishek Jagushte	Cache Hierarchy Management for Address Translation
Sanjeev Kumar	SrcMix: Mixing of Related Source Languages Benefits Extremely Low-resource Machine Translation

Security Operations Center : CC-101

Abstract:

Even today, security remains one of the most overlooked aspects of an organization’s infrastructure despite having the potential to cause shock waves across all operations, if compromised.

The high costs of traditional SIEM solutions make facilitating centralized infrastructure security challenging for smaller organizations. As a result, they still resort to distributed solutions such as Firewalls, antivirus, spam-checks, etc. with minimal event correlation between them.

Our entirely FOSS-based Security Operations Center intends to solve this problem.

We demonstrate our use of open-source tools such as Apache Kafka and the ELK Stack to implement the traditional SOC workflow of data ingestion, transformation, analysis and visualization in a resource-light, modifiable and easily distributable way.

Our solution encompasses alerts across multiple domains concerning most organizations. These domains include web, mail, network, applications, hosts and endpoints using a rule-based approach as well as ML solutions for anomaly detection.

BharatGen : CC-103

Demo 1: Developing an Indic Language LLM for Enhanced AI Accessibility

Abstract:

The scarcity of Indic language representation in the digital landscape poses significant challenges in providing equitable access to AI solutions. Despite Hindi and other Indic languages being spoken by millions, they comprise less than 1% of the content available online, while English dominates with over 51.2% of digital content. To address this imbalance, we propose the development of BharatGen, a foundational language model (LLM) designed to enhance Indic language accessibility.

BharatGen has been trained on diverse datasets, with models like Bharatgen-2.6B and Bharatgen-5B (ongoing training) incorporating both English and Hindi datasets, alongside 13 other Indian languages including Assamese, Bengali, Gujarati, and Tamil. The training involved 8.6 trillion tokens, with 3.6 trillion tokens dedicated to Indic languages, ensuring robust multilingual capabilities.

The initiative aims to foster technological sovereignty through data localization, enhance national security and privacy, and promote cultural preservation. By integrating AI into various sectors such as agriculture, healthcare, and education, BharatGen will drive digital inclusion and innovation. Furthermore, public-private partnerships will be leveraged to ensure sustainable AI self-sufficiency. Through BharatGen, we envision a future where AI serves all linguistic communities, preserving cultural identities while driving technological progress in India and beyond.

Demo 2: Speaker-Adaptive Text-to-Speech for Low-Resource Indian Languages

Abstract:

Generating high-quality speech for unknown speakers and low-resource Indian languages remains a challenge in text-to-speech (TTS) systems. We propose a diffusion-based TTS system with a speaker encoder that extracts speaker embeddings from a small audio sample, conditioning a DDPM decoder for multi-speaker synthesis.

Our system initially faced challenges in zero-shot speaker adaptation and duration modeling. To address this, we introduced an attention-based mechanism that extracts duration features from a 2-second mel spectrogram of another audio sample from the same speaker. This improves prosody by aligning speech timing with speaker characteristics.

To enhance zero-shot generation, we employed classifier-free guidance, enabling speech synthesis for unseen speakers while preserving natural prosody. We trained language-specific speaker-conditioned models on the IndicSUPERB dataset for multiple Indian languages, including Bengali, Gujarati, Hindi, Marathi, Malayalam, Punjabi, Tamil, and Telugu.

Our approach improves speaker adaptation and prosody modeling, making Indian language TTS more natural and inclusive.

BodhiTree & EvalPro : CC-105

Abstract:

Grading assignments in introductory programming courses is very common, especially in large classes. To handle this, autograders are widely used. We use our own platform, BodhiTree, an online learning platform that features its own autograder called EvalPro. In our standard grading scenario, we configure input and output test cases, execute the student’s program, and compare the submitted code’s output with the expected results.

Our platform also provides a rubric-based grading interface for non-compiling code, recognizing that students often make small mistakes in an educational setting. This interface allows graders to evaluate such submissions effortlessly with a single click, and the total marks are calculated automatically. To save time, we introduce an AI-powered autograder, TA-Buddy, which utilizes rubrics to suggest relevant ratings for each criterion. A dedicated team of AI enthusiasts works continuously to improve the model, ensuring that the suggested grades align closely with human evaluations, thereby speeding up the grading process.

Our code editor supports single-file, multiple-file, and advanced configurations that allow restrictions on file names and types. It also includes features like running code without submission and uploading code files.

Exams are an essential part of assessment in education, and our platform makes the entire process smooth and hassle-free. Hosting exams is effortless with our system, allowing instructors to set up and manage exams with ease. Exam Mode enables instructors to restrict exam access based on specific IP ranges, ensuring that only students with approved IP addresses can participate in lab exams. Additionally, the Proctor Dashboard provides instructors with a live view of student activities and full control over the exam process.