Pratik Pokharel
Machine Learning. Computer Vision. Systems.
About
I'm an engineer who enjoys understanding how things work and building things that matter. I graduated in Electronics, Communication & Information Engineering from the Institute of Engineering, Thapathali Campus, and now work as a Project Manager at an early-stage startup, where I help turn ideas into products while staying close to the engineering behind them.
Technology is what I build, but curiosity is what drives me. My interests range from machine learning and backend systems to semiconductors, operating systems, and the architecture behind modern software. I enjoy learning across disciplines and understanding the ideas that connect them.
Away from the screen, you'll usually find me watching films, listening to hip-hop, reading, writing down thoughts, or disappearing into a topic that caught my attention for no particular reason. I like conversations that teach me something new, and I try to bring that same curiosity into everything I build.
How I think
An interactive map of my interests and how they connect. Drag nodes to explore. Hover for details.
Projects
Domino AI →
Hybrid RAG system for financial contagion mapping. Combines vector search with knowledge-graph BFS traversal to answer multi-hop exposure queries that plain retrieval can't resolve. Achieved 1.00 retrieval and entity recall across benchmarks.
Medical Image Report Generation →
CNN-Transformer encoder-decoder that automatically writes diagnostic reports from chest X-rays. Achieved 37.68% BLEU-1, 27.96% ROUGE-L, and 31.50% METEOR on the Indiana University dataset.
Parkinson's Diagnosis from EEG →
Spatial-temporal transformer model that detects Parkinson's Disease from raw EEG brainwave data. Handled complex time-series preprocessing, achieving 78% balanced accuracy and 0.89 ROC AUC.
Sambidhaan 2072 →
RAG system that accurately answers questions about the Constitution of Nepal. Uses Groq's LLaMA 3 70B with LangChain for fast retrieval, served through a Streamlit interface.
Real-Time Handwriting Recognition Pen →
Hardware-software system using Arduino Nano and MPU-6050 sensor to track hand movements. Custom dataset of A-Z motions, classified at 90% accuracy with a CNN-LSTM model.
Multimodal AI for Parkinson's →
Fuses MRI and DAT-SPECT brain scans using ResNet-18 feature extraction fed into a transformer. Multimodal approach for improved diagnostic confidence.
Currently exploring
Agentic AI
Autonomous systems that plan, reason, and execute multi-step tasks — tool use, memory, and self-correction in LLM-driven agents.
Medical AI & Neuroimaging
Diagnostic models combining brain scans, EEG signals, and clinical data for early detection of neurodegenerative conditions.
Computational Neuroscience
How biological neural circuits encode, process, and transmit information — bridging neuroscience and machine learning.
Knowledge Graphs & Hybrid Retrieval
Combining structured graph traversal with vector search for multi-hop reasoning over complex document collections.
Semiconductor Architecture
Understanding chip design, transistor physics, and the hardware constraints that shape modern computing.