4th-year CS (AI/ML) at PES University, Bengaluru.
I work on memory and representation in neural networks — what gets stored, where, and whether it survives. Currently: whether domain knowledge can live in the initial recurrent state of a linear-attention model instead of its weights.
- PAR-VAE: A Physics-Constrained Generative Audit of CT Severity Classification — 1st author, accepted at MIUA 2026 (Springer LNCS)
- High-frequency soil hydrothermal observations from a semi-arid monsoon catchment in southern India, 2016–2025 — 2nd author, Nature Portfolio, under revision · dataset
PAR-VAE — physics-regularized VAE for CT imaging Ties an 85-dim latent space to 14 interpretable physics attributes (density, geometry, texture). Studies the interpretability/accuracy tradeoff, and finds that much of the remaining gap to a black-box CNN at milder disease stages appears data-intrinsic rather than a modelling limitation.
GREM — distilling multi-agent LLM reasoning into a small re-ranker Gemini agents generate reasoning about multi-hop retrieval failures during training; verified chains go to MongoDB Atlas; a 22M cross-encoder learns from them and runs with no LLM calls at inference. Built for the Google Cloud Rapid Agent Hackathon. README documents what the evaluation does and doesn't establish.
KG-EVOLVE — episodic memory over frozen KG embeddings Multi-agent reasoning on top of frozen ComplEx embeddings, with writeback gated on validation. Found a failure mode where early low-quality memories skew later reasoning.
BGSAVM — graph-based patch selection for 3D point clouds Uses evidential uncertainty and motif participation to choose patches, with a ViG backbone over unstructured 3D geometry.
PyTorch · PyTorch Geometric · HuggingFace · NumPy · Python