M.S. by Research Candidate · Roll: 24M2133 CFILT Lab, IIT Bombay · Advised by Prof. Pushpak Bhattacharyya

Vijendra Kumar Vaishya

AI Researcher & Systems Engineer · Computer Science & Engineering, IIT Bombay

I am a graduate researcher in Computer Science at IIT Bombay working across the machine learning stack—from foundational model design and representation alignment to low-level systems execution.

My research focuses on making language and vision systems both reliable and efficient:

Research & Thesis Publications

Accepted at WACV 2027

MIST: Mamba-based Integrated Spatio-Temporal Model for Intracranial Vessel Segmentation in DSA Sequences

  • Developed a dual-branch spatio-temporal architecture for DSA vessel segmentation using Mamba-based State-Space Models (SSMs).
  • Designed a Cross-Attention Fusion (CrossSTF) module integrating MinIP-derived spatial priors with multi-scale temporal features, refined using topology-preserving deep supervision.
  • Achieved a Dice score of 0.7801 on the independent DSCA benchmark without any retraining or fine-tuning, demonstrating robust cross-dataset generalization.
EACL 2027 (Under Review)

Beyond Single-Axis Fairness: Learning to Detect Intersectional Biases

  • Built a retrieval-based framework for intersectional bias detection, recast as retrieval over classification.
  • Introduced BiasRetriever, a contrastively trained retriever using canonical single-axis stereotype embeddings.
  • Boosted Jaccard score by 18 points over the strongest LLM baseline across disjoint corpora.
KCDH Lab, IIT Bombay · In collab with PHFI

Ontology-Aware Persona-based Chatbot for Tobacco Cessation

  • Engineered a 2-stage conversational architecture utilizing a decision-tree knowledge graph to dynamically classify patient readiness across 5 cessation stages and assess Fagerström dependence severity.
  • Built a 6,900-concept clinical ontology with 33k+ edges and designed an ontology-aware text segmenter, reducing broken mid-sentence chunk splits from 84.6% to 5.6% compared to fixed-size chunking.
  • Achieved 58% Recall@3 (29/50) and 0.497 MRR using ontology-aware chunking across 50 clinical queries, outperforming baseline semantic approaches.
MeitY, Govt. of India · NLTM-Bhashini

Bhashini Speech-to-Speech Machine Translation (SSMT)

  • Architected a full-stack annotation platform and transformer labeling model for Marathi Multi-Word Expression detection on 4,260 manual annotations to accelerate low-resource language corpus curation.
  • Engineered an automated MLOps retraining engine, expanding dataset size from 4,260 to 13,900+ annotations.
  • Achieved 85.05% test F1 and 90.76% Complex Predicate detection, surpassing baseline performance.
Pre-print

Subset Selection for Fine-Tuning: A Utility-Diversity Balanced Approach for Mathematical Domain Adaptation

Proposed a utility-diversity subset selection framework tailored for mathematical domain adaptation in large language models.

Systems & Self Projects

Bare-Metal C++17 · Apple Silicon / ARM64

Fused SwiGLU / Transformer FFN Inference Engine

  • Built bare-metal C++17 engine for LLaMA/Mistral FFN with a 5-stage optimization ladder to INT8.
  • Hand-wrote 8-wide ARM NEON microkernel, reaching 37–84x speedup over scalar baseline.
  • Applied INT8 quantization achieving 41.7 GFLOP/s throughput at <0.6% NRMSE error.
  • Fused activation and quantization ops in registers, exploiting P-/E-core scheduling for 1.43–1.48x further gain.
SwiGLU Microkernel Throughput Ladder
1. Baseline Scalar C++ (O3) 0.50 GFLOP/s (1.0×)
2. Auto-Vectorized Clang NEON 18.6 GFLOP/s (37.2×)
3. Hand-Written 8-Wide NEON Microkernel 36.8 GFLOP/s (73.6×)
4. Fully Register-Fused NEON + INT8 Quant (Ours) 41.7 GFLOP/s (84.2× Speedup)
Machine Learning Theory · NumPy

Custom Gaussian Process & Bayesian Optimization Engine

  • Implemented from-scratch GP surrogate using Cholesky factorizations ($LL^T$) and log-marginal likelihood optimization.
  • Formulated Expected Improvement (EI) & Probability of Improvement (PI) acquisition functions via analytical Gaussian CDF/PDF, converging in 50 evaluations.
Full-Stack & Scalable Platform

StoriesbySafarnama — Collaborative Digital Storytelling Platform

  • Founded and built Safarnama, a private platform for creating and sharing collaborative digital memory timelines.
  • Designed and deployed full-stack architecture for timeline creation and collaborative editing, integrating media uploads and secure sharing. Built scalable backend to manage user data, timeline content, and media assets.
  • Integrated DODO Payments for secure checkout, transaction tracking, and automated refund workflows.

Interactive Research Terminal

Type commands or click quick pills to inspect publications, C++17 NEON benchmarks, and academic stats in real time.

📄 CV
vijendra@iitb:~$
NORMAL main* IIT Bombay CSE · CFILT Lab 14 cmds
UTF-8 zsh 5.9 06:00 Click pills or type commands

Graduate Coursework Projects @ IIT Bombay

CS 725 · Fall 2024

Medical Image Classification Using Vision Models

Instructor: Prof. Preethi Jyothi
  • Built an AutoML optimized Vision-Transformer pipeline for multi-modal medical image classification.
  • Exhibited 6.7% average gain in Accuracy and AUC on multiple MedMNIST datasets over ResNet and DenseNet.
CS 626 · Fall 2024

Multi-Hop Question-Answering & Reasoning

Instructor: Prof. Pushpak Bhattacharyya
  • Implemented self-prompted chain-of-thought with 6 reasoning paths and studied extensibility to Indic tasks.
  • Infused LightRAG with GPT-4o, increasing EM by 2% on HotpotQA and 2Wiki compared to Zero-Shot CoT.
CS 769 · Spring 2025

Subset Selection for Fine-Tuning LLMs

Guide: Prof. Ganesh Ramakrishnan
  • Proposed a utility-diversity subset selection framework using a Chain-of-Thought metric combining confidence divergence, backtrack frequency, and complexity.
  • Achieved 50% accuracy at 1900 examples, outperforming DPP and random selection.
CS 772 · Fall 2025

Diagram Visual Question Generation (LeJEPA + BLIP-2)

Instructor: Prof. Pushpak Bhattacharyya
  • Engineered a two-stage VQG pipeline with LeJEPA pre-training and BLIP-2 Q-Former fine-tuning (+4 BLEU-4 points).
  • Surpassed 7.8B SOTA model on DiagramQG in Bert-Score (93.55 vs 90.98) with LeJEPA + FLAN-T5-Large (5.5x parameter reduction).
CS 791 · Fall 2025

Controlled LLM Decoding via Sequential Sampling

Instructor: Prof. Sunita Sarawagi
  • Evaluated 7 controlled decoding strategies including SMC and TSMC, achieving a 146% increase in target rarity reward over greedy decoding.
  • Demonstrated that SMC achieves 32% higher sample diversity than SIS while eliminating particle weight degeneracy.
Completed Graduate Coursework: Foundations of Machine Learning (CS 725) Deep Learning for NLP (CS 772) Speech & NLP and the Web (CS 626) Optimization in ML (CS 769) Probabilistic Foundations of AI (CS 791) Organization of Web Information Software Lab

Academic Record & Education

Indian Institute of Technology Bombay (IIT Bombay)

2024 – 2027
M.S. by Research in Computer Science and Engineering · CPI: 8.9 / 10

Roll Number: 24M2133. Core Coursework Credits: 28 | Total Credits: 58.
Advised by Prof. Pushpak Bhattacharyya in the CFILT Lab. Research focus: Mamba State-Space Models for 3D spatio-temporal segmentation, intersectional bias evaluation in LLM embedding spaces, clinical ontology RAG architectures, and bare-metal ARM NEON C++17 optimization.

Jabalpur Engineering College (RGPV)

2020 – 2024
Bachelor of Technology (B.Tech) in Information Technology · CPI: 8.06 / 10

Coursework in Algorithms, Operating Systems, Computer Architecture, Compiler Design, Distributed Computing. Solved 500+ algorithmic problems on LeetCode across dynamic programming, trees, and graph algorithms.

DPS Nigahi (CBSE Class XII)

2020
Senior Secondary Certificate Examination · 96.00% Aggregate

Secured 3rd Rank across the entire district with 96.00% aggregate score. Achieved a perfect 100% score in Mathematics. Qualified the JEE Advanced examination.

DPS Nigahi (CBSE Class X)

2018
Secondary School Examination · 92.40% Aggregate

Secured 92.40% aggregate in CBSE 10th Board Examinations.

Technical Skills

Programming Languages

Python C++ (17) C Bash Scripting

ML / DL Frameworks

PyTorch Hugging Face Transformers OpenCV NumPy CUDA

ML Techniques & Architectures

Mamba / State-Space Models (SSMs) Vision Transformers Retrieval-Augmented Generation (RAG) Contrastive Learning LLM Fine-Tuning & INT8 Quantization Bayesian Optimization

Systems & Backend

ARM NEON SIMD Flask MLOps Pipelines FAISS Docker MySQL

Developer Tools

Git LaTeX Postman

Positions of Responsibility

  • Student Companion | ISCP, IIT Bombay: Mentored 5 students in academics, social, and extracurricular activities for holistic well-being.
  • Technical Team Member | AACL 2025, IIT Bombay: Supported technical execution of 5-day international conference.

Achievements & Hobbies

  • 3rd District Rank in Class 12th CBSE Board Examinations; qualified JEE Advanced.
  • 100% in Mathematics in CBSE Boards (2018); solved 500+ problems on LeetCode.
  • Hobbies: Gym, Reading, Badminton, Cooking, Cricket, Expressive Drawing, Chess & learning to play the Guitar.

Get in Touch

Open to AI research collaborations, systems engineering roles, graduate fellowships, and academic inquiries.

📧
Direct Email vijendra9753@gmail.com

🏛️ Research Office: Department of Computer Science & Engineering, IIT Bombay, Powai, Mumbai 400076, India