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.

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.

Reproduced three papers spanning LLM evaluation, agent behavior and cross-view geolocation.

Who Can We Trust? LLM-as-a-Jury for Comparative Assessment

Introduces BT-σ, a judge-aware extension of the Bradley–Terry model that jointly infers item rankings and per-judge reliability from pairwise LLM comparisons. Addresses the problem that LLM judges vary in accuracy and consistency rather than being equally trustworthy.

View Reproduction

Must All Negatives Be Pushed Away Equally? Uncertainty-Aware Cross-View Geo-Localization

Models per-sample uncertainty with a Normal Inverse Gamma distribution to weight negative pairs unevenly during contrastive training, instead of pushing every negative away with equal force. Aims to improve matching between ground-level and satellite imagery for geo-localization.

View Reproduction

When Does Predictive Inverse Dynamics Outperform Behavior Cloning?

Compares Predictive Inverse Dynamics Models (PIDM) against standard Behavior Cloning (BC) for imitation learning, showing PIDMs trade added prediction bias for reduced variance. Finds PIDMs reach comparable performance with far fewer demonstrations than BC in navigation tasks.

View Reproduction

I built a cross-view localization system that determines a UAV's position by matching its onboard imagery against a satellite reference database, enabling navigation in GPS-denied environments. The pipeline embeds both UAV and satellite imagery using Vision Transformers and retrieves the closest matching reference tile with an IVF (Inverted File Index) index for fast approximate nearest-neighbor search over large-scale satellite coverage, reaching 92% Recall@1 — correctly identifying the true location as the top candidate in the vast majority of queries.

I extended the system beyond visible-spectrum imagery to also match against thermal imagery, letting the UAV localize itself in low-visibility conditions such as night operations, fog, or smoke where standard optical cameras fail. To make the system viable for onboard deployment, I optimized the inference pipeline with model quantization and TensorRT FP16 precision, accelerating scene-matching time by 15× while staying within 1% of baseline accuracy, bringing the full pipeline within real-time constraints on the embedded flight system.