🎓 Final-year B.E. student in Electronics, Communication and Information Engineering at the Institute of Engineering, Tribhuvan University (IOE Thapathali Campus). I build across the ML stack — from deep learning for medical imaging, speech processing, and graph neural networks to pushing inference onto resource-constrained edge devices.
Co-author of IsoNet (multimodal target speech extraction), 1st-place winner at KU Hackfeast 2025, and President of Tensor IOE. Always happy to talk AI, ML, and edge computing.
- 🔭 Currently working on spatiotemporal graph transformer for early Alzheimer's detection.
- 🌱 Learning edge AI deployment and trustworthy / interpretable ML.
- 💬 Ask me about AI, deep learning, and edge computing.
🎯 IsoNet — Multimodal Audio-Visual Target Speech Extraction(1st place, KU Hackfeast 2025 · paper accepted at JIEE)
Solves the cocktail-party problem using a 4-mic array + face-conditioned visual embeddings in a U-Net, achieving 9.31 dB SI-SDR and beating oracle beamformers. (repo coming soon)🚶 Human Presence Detection — Distributed Edge Inference🔗
Tile-based CV on ESP32-CAM (Sobel edge density) + LBCNN on ESP32-S3 over ESP-NOW for real-time distributed presence detection (82% acc).🤖 Detection of Hallucinations in LLM Reasoning Chains🔗
Lightweight classifiers (embedding-similarity → LSTM + attention) for flagging hallucinations in step-by-step LLM reasoning.🧠 Spatiotemporal Graph Transformer for Early Alzheimer's Detection(ongoing FYP)
Graph Attention + Temporal Transformer on rs-fMRI (ADNI) with self-supervised contrastive pretraining and Integrated-Gradients interpretability.
- 💼 LinkedIn: @i-r-pokharel/
- 📄 Résumé: @ishwor
- 🌐 GitHub: @Ishwor-git