Skip to content

Projects

GraphBit

530+ stars

Open Source · 530+ Stars

An agentic AI framework featuring a dual-layer memory architecture for context-aware autonomous agents. Combines SQLite-backed persistent memory with an in-memory transient layer for fast, session-scoped working memory.

Key Contributions

  • Designed and implemented the memory mechanism as core developer
  • Architected a dual-layer memory system for context-aware agent behaviour
  • Built a persistent memory layer backed by SQLite for long-term knowledge retention and retrieval across agent sessions
  • Developed an in-memory storage layer for transient, session-scoped context, enabling low-latency access to short-lived working memory during agent reasoning
  • Enabled agents to maintain coherent state across sessions while keeping real-time reasoning latency low

Dual-GAT Fault Diagnosis

Scientific Reports (Nature), 2025

Developed a dual graph attention network (Dual-GAT) for automated fault detection and classification in photovoltaic inverter systems. The model leverages both spatial and spectral graph convolutions to capture complex inter-feature relationships, enabling robust diagnosis under noisy and imbalanced operating conditions.

Key Contributions

  • Proposed a Dual-GAT architecture combining spatial and spectral graph convolutions
  • Captured complex inter-feature relationships in photovoltaic inverter fault signals
  • Achieved robust fault diagnosis under noisy and imbalanced conditions
  • Published in Scientific Reports (Nature), vol. 15, no. 1, p. 31330, 2025

Hyperspectral Image Classification

ICCIT 2022 (IEEE)

Proposed an attention-aware memory aggregation network that combines spectral-spatial feature extraction with an external memory module for improved classification of hyperspectral remote sensing imagery.

Key Contributions

  • Designed an attention-aware memory aggregation network for hyperspectral image classification
  • Combined spectral-spatial feature extraction with an external memory module
  • Developed an attention mechanism that selectively emphasises discriminative spectral bands
  • Memory module retains representative class prototypes across training iterations
  • Published at the 25th ICCIT, IEEE, 2022