Skip to content
View aaryaupadhya12's full-sized avatar

Block or report aaryaupadhya12

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
aaryaupadhya12/README.md

Aarya Upadhya

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.


Publications

  • 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

Projects

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.


Stack

PyTorch · PyTorch Geometric · HuggingFace · NumPy · Python


aarya.upadhya@gmail.com

Pinned Loading

  1. Physics-Attribute-Regularized-VAE Physics-Attribute-Regularized-VAE Public

    Class Overlap Analysis using Physics Attribute regularized Auto Encoder

    Jupyter Notebook 1 1

  2. KG-EVOLVE KG-EVOLVE Public

    Training knowledge graph embeddings with PyKEEN and experimenting with agent-driven knowledge graph schema evolution.

    Jupyter Notebook 1 1

  3. GREM GREM Public

    Episodic Distillation for Verified RAG using Google Cloud + MongoDB Atlas

    Python 1

  4. PeakShift PeakShift Public

    At retail stores, management staff offer to work in the stores during very busy times like weekends and holidays. This is an app which faciliitates this initiative

    Python 1

  5. GRPO_Reasoning_PostTraining GRPO_Reasoning_PostTraining Public

    Post Training on LLM for reasoning with Group Relative Policy Optimization

  6. MultiModal_Concept_Bottleneck_Models_Aufit MultiModal_Concept_Bottleneck_Models_Aufit Public

    MultiModal_Concept_Bottleneck_Models_Aufit