Research

Research project under the Indian Space Research Organisation focused on cross-modal image registration of lunar terrain data for autonomous descent navigation, targeting onboard FPGA deployment.

Working on the problem of cross-modal alignment between LiDAR intensity images, Digital Elevation Models (DEM), and optical imagery. The focus is on structural and geometric differences across these modalities and how they affect feature matching and representation. Structural and geometric representations are used as the primary modality bridge to reduce radiometric dependence across sensors.

Benchmarked hybrid registration pipelines across multiple lunar terrain regions using LOLA DEM and TMC-2 orthoimage pairs. Presented progress at SAC-ISRO, Ahmedabad.

Industry role focused on deploying machine learning systems for real-world applications, particularly in speech and language processing.

Worked on integrating speech-to-text systems with downstream NLP pipelines, focusing on robustness under noisy real-world data. Contributed to backend ML services and inference pipelines with emphasis on evaluation, reproducibility, and system stability.

Early-stage experimentation and prototyping in speech processing systems. Transitioned into full-time role.

Multimodal pipeline combining speech input, document ingestion, and retrieval-based generation within a single end-to-end system.

Retrieval-augmented generation system for structured financial document question answering, with domain adaptation via fine-tuned Flan-T5.