Research

My research is on LLM-based decision support for autonomous driving, with a focus on safety, constraint compliance, and evaluation. I build retrieval-augmented and agentic LLM pipelines, and the evaluation harnesses that measure when they can be trusted. Earlier, I worked on NLP and neuro-inspired models for time-series classification.

LLM Advisory Agents for Autonomous Driving Decision Support (In progress)

I am building LLM-based advisory agents that support driving decisions in simulated traffic scenarios, together with an evaluation framework that measures whether their advice is safe, follows traffic rules, and is well-formed. The work combines retrieval over traffic rules, structured (schema-validated) outputs, and scenario-based safety evaluation.

Hybrid Spiking Neural Network -- Transformer Video Classification Model

Inspired by hybrid Transformer-CNN models, we design a Neuro-inspired hybrid Transformer-SNN model. It closely mimics the cortical column structure of the brain, which lets it capture temporal information as well as spatial information. The model is implemented and tested in text and video time-series classification. Part of this work became my BSc Thesis.

Projects

I have also worked on a variety of personal and course projects, some of which are featured below. For a full list, see my GitHub page.

NetOps GenAI Assistant: Agentic RAG for Incident Triage

An agentic RAG assistant for network-incident triage. It retrieves runbook documentation, pulls the affected site’s KPI metrics through a tool call, and returns a structured decision (severity, likely cause, recommended steps, cited sources) through a FastAPI service. I evaluated it with 6 local LLMs and 4 test suites (LLM-judge and ROUGE-L/BLEU): retrieval raised correctness by +39.6 pts on facts that cannot be known without retrieval (6 of 6 models), and model choice decided whether retrieved context was trusted over the model’s prior.

Test-Time Adaptation for Polyp Segmentation under Domain and Task Shift

I reimplemented NA-SegFormer (Transformer encoder, CNN decoder) and benchmarked five test-time adaptation (TTA) methods on three Kvasir datasets under in-distribution, domain-shift, and task-shift conditions, including a new variant I designed (CF-Geo-VPTTA). VP-TTA was the only method that improved every dataset, with up to +17 Dice pts under combined domain and task shift, while augmentation-based methods helped in-distribution but hurt under shift.

AddressBook: Spring Boot REST Service with CI/CD

A Java 21 / Spring Boot 3 web application with REST endpoints and HTML views for managing address books, persisted through Spring Data JPA. It has JUnit unit tests and end-to-end API integration tests, and a GitHub Actions CI/CD pipeline that built, tested, and deployed the app to Azure App Service.

Study on Effects of UNet's Variations in Polyp Segmentation

I studied how architectural changes to U-Net affect polyp segmentation, with a focus on how the depth of the U-Net changes its performance.

Pacman Playing Agent

I developed a Q-Learning agent that learns and plays the Pacman game.

Implementation of Spiking Neural Architectures

Implemented several Neural Architectures as prerequisites to my BSc Project. The list is as follows: (Image Credit: Wikimedia)

1. Neural Models: LIF, ELIF, AELIF
2. Encoding Methods: Poisson, TTFS, Positional
3. Currents: Sinusoidal, Constant, Step Current, Multi Step Current

Implemented in PyTorch.

Comprehensive Analysis of Information Retrieval Algorithms

Comprehensive implementations of several Information Retrieval Systems. The list is as follows:

1. WebIR with boolean queries
2. WebIR with wildcard query (Trie, and Permuterm), and spellchecking (Soundex, and Levenshtein Distance)
3. WebIR with Blocked Sort-Based Indexing (BSBI)
4. Ranked WebIR with sorting queries based on relevance (Okapi BM25, Language Model, and TF-IDF)
5. Comprehensive Performance Comparison of IR methods

Corridor Game

Using C++17, I developed a client-server version of the Quoridor board game for 2-4 players. Clients and server communicate over HTTP with JSON messages, and the server validates every move, including a reachability check on wall placements. Unit-tested with Catch2. (Image credit: Wikimedia)

Earlier Projects

• Loss Function Optimization with Particle Swarm Optimization (2024): particle swarm optimization for loss-function minimization.
• Review of Ant Colony Methods in SCP (2022): comparison of Ant Colony System methods for the Set Covering Problem.
• Simulated Annealing for CSP (2022): simulated annealing for the Cutting Stock Problem.
• Genetic Method for N-Queen Problem (2022): genetic algorithm for the N-Queens problem.
• ClassBoom (2021): Django e-learning platform with class recordings, homework, and grading.

Teaching & Academic Service

I have also been an active member of our academic community. Some of my main contributions are as follows.

Teaching Assistantship

• TAed for: SYSC 4806 (Software Engineering Lab) (two terms) and SYSC 4130 (Human-Computer Interaction).
• Ran labs and project presentation sessions for ~130 students (SYSC 4806) and Problem Analysis sessions for ~92 students (SYSC 4130).
• Helped students design and implement their software projects and UI prototypes, and supported them via office hours and as Discord moderator for SYSC 4806.
• Graded lab checkouts, recorded topic presentations, and lab submissions (SYSC 4806), and midterm/final exams (SYSC 4130).

Teaching Assistantship

• TAed for: Design and Analysis of Algorithms, Graph Theory, Logical Circuits & Architecture, Combinatorics, Calculus I, and Basic Programming.
• Held classes for 250+ students and prepared 25+ assignment series.
• Designed real-world problems and aggregated widely used problems from academic references.
• Graded 1000+ student submissions, quizzes and exams.

Elected Member of the Computer Science Students’ Scientific Chapter

• Organized workshops on career development, resume writing, and programming with distinguished speakers.
• Organized orientation for the new students on various topics such as programming languages/disciplines.

Design and source code partially borrowed from Jon Barron's website