Nishka Katoch

Addressing the critical challenge of limited annotated data in biomedical imaging, I developed a microscopy pseudo-labeling pipeline using unsupervised image translation techniques. The system generated high-quality pseudo labels by translating between labeled and unlabeled domains without requiring prior annotation in the target domain. By implementing a SyncDiff-based domain adaptation pipeline with VQ-VAE compression, I improved segmentation accuracy on out-of-distribution microscopy data by 12% while accelerating training speeds and maintaining 95% image fidelity.

The method was validated across three increasingly dissimilar image domains, achieving a mean Dice score of 0.736 ± 0.005 on Scanning Electron Microscopy (SEM) datasets using synthetic data translated from Transmission Electron Microscopy (TEM). This approach significantly accelerated the annotation process by providing researchers with refined starting points for manual labeling, bridging the gap between different microscopy modalities in biomedical research.

To address data scarcity in neurodegenerative disease research, I developed a privacy-preserving image translation framework using latent diffusion models for histological image analysis. The system converted source images into latent encodings and reconstructed them across different modalities using guided diffusion models, with segmentation masks integrated directly into the latent space. This architecture enabled domain adaptation without requiring access to real patient data, ensuring compliance with medical privacy regulations.

The framework achieved an 80% Dice score using synthetic contrast data, demonstrating effective segmentation performance on neurological tissue across diverse electron microscopy modalities. By minimizing the number of translation models required and operating entirely on latent encodings, this approach offered a scalable solution for studying Alzheimer's disease and multiple sclerosis while maintaining strict data privacy standards.

I conducted a comprehensive empirical study on the scaling properties of speech recognition and natural language understanding models in multispeaker settings. The research examined CTC-based speech recognition with pretrained wav2vec2 encoders and a bimodal NLU intent classification model, comparing different audio preprocessing approaches including raw features, RAVE variational autoencoder latents, and wav2vec2 acoustic representations. Through systematic benchmarking, I derived predictive scaling laws: power laws for ASR and audio accuracy, and logarithmic laws for text-to-audio embedding similarity.

The findings revealed that wav2vec2-based models followed predictable scaling patterns with optimal performance at specific model widths and layer counts. Pretrained acoustic models demonstrated superior scaling compared to training from scratch, with wav2vec2 offering the greatest advantage. These insights directly informed efficient deployment strategies for multilingual ASR systems in low-resource environments, helping organizations optimize model selection based on available data and computational constraints.

I led a comprehensive analysis of single-nucleus transcriptomes from the prefrontal cortex of 48 individuals to understand the cellular context of Alzheimer's disease pathology. Using six different dimensionality reduction and visualization methods—including Truncated TSVD, kernel PCA, Isomaps, t-SNE, UMAP, and PHATE—I revealed distinct geometrical structures in cell-type distributions and identified fine-grain patterns correlating with AD progression markers (CERAD scores and Braak stages).

The analysis uncovered cell-type-specific pathological associations, identifying Mic1 as an AD-associated subpopulation and Oli1 as a no-pathology subpopulation. Visualizations revealed a gender bias in AD diagnosis prevalence among females, aligning with findings in existing literature. This multi-scale approach to genetic and clinical feature exploration provides a foundation for identifying AD-pathology-associated cell populations and potential biomarkers for early disease detection.

Developed a novel hybrid classical-quantum architecture for detecting AI-generated synthetic images by leveraging Parameterized Quantum Circuits (PQC) for binary classification. The system combines traditional deep learning feature extraction with quantum neural networks to identify manipulated content from contemporary generative models including Flux, DALL-E, and Gemini.

Implemented a multi-stage pipeline featuring Class Attention Vision Transformer (CaiT) for feature extraction, reducing high-dimensional image data to 64-dimensional feature vectors suitable for quantum processing. Designed and evaluated quantum circuits with 4 and 8 qubits using Hadamard gates for superposition, Rotation-Y gates for parameterization, and CNOT gates for entanglement. Explored two architectural approaches: a linear 8-qubit circuit duplicating the 4-qubit structure, and a hierarchical circuit with cross-entanglement distinguishing between coarse and fine feature groups for multi-scale learning.

The research addressed critical challenges at the classical-quantum interface, including dimensionality reduction, quantum encoding strategies, and circuit depth optimization. This work demonstrates the potential of quantum computing to enhance pattern recognition in deepfake detection while identifying practical deployment considerations for hybrid quantum-classical systems in computer vision applications across financial services, identity verification, e-commerce, and legal compliance.

Led the technical design and research architecture for a comprehensive benchmarking framework evaluating embedding generation and vector quantization techniques for Tech Mahindra's SharePoint Retrieval-Augmented Generation (RAG) system. Coordinated a team of interns to develop a Python-based pipeline addressing performance and scalability challenges in enterprise document retrieval systems.

Designed a modular five-component architecture encompassing PDF document processing with PyMuPDF, embedding generation using the Nomic model (768-dimensional vectors), vector database operations with Milvus Lite, Product Quantization benchmarking with FAISS, and comprehensive reporting analytics. The system was architected to achieve 8-32x compression ratios while maintaining search accuracy through systematic evaluation of quantization parameters (16 subvectors, 4-8 bits per subvector).

Conducted extensive research on vector quantization methods, synthesizing technical knowledge from foundational papers on Product Quantization and modern similarity search approaches. Established technology stack decisions, distributed team responsibilities, and provided technical mentorship on embedding models, vector databases, and accuracy-storage trade-offs. Created detailed architectural specifications and experimental workflow design covering initialization, document processing, embedding generation, vector storage, benchmarking, and results reporting phases.

I developed an end-to-end computer vision pipeline for automated interior design generation, leveraging the Segment Anything Model (SAM) for precise scene segmentation combined with diffusion models to create diverse design variations from raw house images. The system processed input photographs, identified structural and decorative elements, and generated multiple stylistically coherent design alternatives while maintaining spatial constraints.

I designed a text-to-image control interface that enabled natural language-based customization of design outputs, allowing users to specify style preferences, color schemes, and furniture arrangements through conversational prompts. This project bridged the gap between AI capabilities and practical interior design applications, making professional-quality design exploration accessible to homeowners and designers.

I designed and deployed a production-ready Hindi Automatic Speech Recognition (ASR) system achieving over 90% accuracy across diverse regional datasets and dialectal variations. Through systematic fine-tuning of Indic language models, I improved robustness for multilingual speech-to-text tasks, addressing the unique phonetic and linguistic challenges of Hindi and its dialects.

Conducted an extensive comparative analysis of 20+ ASR and transformer-based architectures, evaluating performance trade-offs across accuracy, latency, and computational requirements. This research informed strategic decisions on model selection and optimization approaches, establishing best practices for deploying ASR systems in low-resource language environments.

I developed four critical modules for a Django-based web application that digitized manual workflows at the Centre for Railway Information Systems, increasing operational efficiency by 70%. The modules streamlined data entry, processing, and reporting procedures that previously required significant manual intervention.

Through database optimization efforts, I reduced redundancy by 60%, significantly enhancing system scalability and improving query performance. Additionally, I mentored junior developers on software engineering best practices, establishing code review processes and documentation standards that improved team productivity and code quality across the organization.

I investigated predator-prey behavioral dynamics in UAV swarms, exploring how multi-agent systems could coordinate in competitive scenarios. I built a ROS-Unity integration achieving sub-80ms latency for real-time VR simulations, enabling researchers to visualize and interact with swarm behaviors in immersive environments.

The simulation framework modeled emergent behaviors in multi-agent collaboration, providing insights into coordination strategies, pursuit-evasion tactics, and collective decision-making. This research had applications in autonomous drone systems for search-and-rescue operations, surveillance, and adversarial robotics scenarios where multiple agents must adapt to dynamic, competitive environments.

Submitted to HRI 2020, a ACM/IEEE International Conference on Human-Robot Interaction.