I build practical products with AI/ML, full-stack web development, and developer-focused tools.
Continual learning architecture that restructures itself to prevent catastrophic forgetting. Heads split and merge dynamically, pathways self-modulate, and slow-pathway weights are selectively consolidated.
- Novel mechanisms: Morphogenic attention (MA), plasticity-gated MLP (PG-MLP), architecture genome vector (AGV), cognitive budget allocator (CBA)
- Results: 98.5% accuracy on Split-MNIST with 0.86% forgetting (vs EWC: 97.35% / 2.29% forgetting)
- Approach: Architecture itself adapts proactively, not reactive weight consolidation
- Comprehensive benchmarks, multi-seed evaluation, publication-quality ablations
- Tech: Python, PyTorch, research-grade continual learning
- Status: Research prototype with full paper and reproducible experiments
Platform connecting students directly with mentors for college admissions guidance. Built for the 1.3M+ students scattered across Reddit, Discord, and anonymous forums looking for structured, real mentorship.
- Full mentor discovery engine with ML-based matching
- Bookable sessions, direct messaging, in-app video via Agora
- Admin moderation, Stripe payments, university-specific feeds
- Tech: Next.js 16, TypeScript, PostgreSQL, Clerk, Stripe, AWS S3
- Live: linkU demo
Decision support tool for farm-level agronomy. Structured assistance from field data and optional leaf imagery with explainable outputs, full audit trail, and transparent uncertainty boundaries.
- Crop suitability & fertilizer recommendations from soil nutrients and weather
- Lightweight ONNX leaf disease model (~50MB)
- Full REST API + SQLite history with CSV export
- Built for technical evaluators and ML engineers exploring agriculture verticals
- Tech: Flask, scikit-learn, PyTorch, ONNX, Vercel
- Live: fieldsense-ai-platform.vercel.app
Continual learning framework that expands model capacity only when learning dynamics indicate it's necessary. Self-regulating system combining novelty signals, meta-parameters, and gated knowledge memory.
- Prevents catastrophic forgetting while maintaining efficiency
- Adaptive capacity grows on demand, not worst-case assumptions
- Combines meta-learning with capacity expansion
- Research prototype addressing core challenge in continuous autonomous systems
- Tech: Python, PyTorch, meta-learning research
Languages: Python, TypeScript, JavaScript, Go, Rust
Web: Next.js, React, Node.js, Flask
Data & ML: PyTorch, scikit-learn, TensorFlow, ONNX
Infrastructure: Docker, Kubernetes, Vercel, AWS
📧 poudeldarshan44@gmail.com
🔗 github.com/rsd-darshan
🌐 darshanpoudel.netlify.app