I build practical products with AI/ML, full-stack web development, and developer-focused tools.
A novel continual learning architecture that addresses catastrophic forgetting through dynamic adaptive mechanisms. The system achieves 98.5% accuracy on Split-MNIST with only 0.86% forgetting rate, significantly outperforming EWC (developed by DeepMind, Google). Validated with comprehensive multi-seed benchmarking and detailed ablation studies to isolate the contribution of each component. Full research paper and reproducible implementation available.
Tech: Python, PyTorch | Status: Research prototype with full paper
A cross-platform AI sidebar that operates continuously in the background, providing real-time contextual assistance by analyzing your current screen and work context. Built with Electron for seamless desktop integration, React for a responsive and intuitive interface, and FastAPI backend for intelligent processing. The system seamlessly understands the application, document, or task you're currently working on, enabling context-aware suggestions, content summaries, and intelligent assistance without requiring manual input or application switching. Perfect for developers, writers, researchers, and knowledge workers who need always-on contextual support while maintaining workflow continuity.
Tech: Python, Electron, React, FastAPI | Status: Active development
An intelligent agricultural decision-support platform that leverages machine learning to optimize farming outcomes at scale. Features include data-driven crop recommendations based on soil conditions, weather patterns, and historical yield data, fertilizer optimization algorithms that reduce costs while maintaining productivity, and computer vision-based leaf disease detection for early intervention and disease management. In simple language: get personalized crop suggestions for your farm, optimize fertilizer usage to cut costs, and diagnose plant diseases just by uploading a photo. Exposes all capabilities through a production-ready REST API with explainable AI outputs showing the reasoning behind recommendations, complete audit trail for regulatory compliance, and persistent history tracking for continuous improvement. Currently deployed with live demonstrations.
Tech: Flask, scikit-learn, PyTorch, ONNX | Status: Live demo deployed
A self-regulating continual learning framework that autonomously expands neural network capacity only when necessary, preventing both catastrophic forgetting and unnecessary over-parameterization. Uses novelty detection signals to identify when existing model capacity is insufficient for new tasks. Employs learnable meta-parameters to dynamically control capacity growth. Integrates gated knowledge memory mechanisms to selectively consolidate and retrieve learned information. Particularly effective for scenarios with streaming data and non-stationary task distributions. Includes research-grade implementation with full reproducibility and comprehensive benchmarking.
Tech: Python, PyTorch | Status: Research implementation
- LinkU — Tried to fix college admissions chaos by connecting students with real mentors instead of Reddit threads
- FunFusion — Made 30 games that fit in 2MB because why not
- MeroKitab — Built a marketplace for buying/selling used books (Nepal focus)
- GharBari — Online grocery store for Nepal, tried to make affordable shopping easier
- RoutRover — Real-time bus tracking so people knew when their bus was actually coming
- MediaScout — Just a tool to find free stock photos and videos without paying
- AlgoLearn — Thought I could teach computer science better than tutorials out there
- RepoExplorer — Wanted an easier way to find GitHub repos with actual filters
- Template-Explorer — Browser for website templates, basically that's it
- RepoGenie — AI helper to set up and optimize projects (probably overly ambitious)
All archived now, but they taught me how to actually build things.
📧 poudeldarshan44@gmail.com | 🔗 github.com/rsd-darshan | 🌐 darshanpoudel.netlify.app