Python Programming
Syntax · OOP · Libraries
BeginnerEngineer. Educator. Always a student.
Good ideas deserve
to be understood.
And built.
I’m Aayush Raj Regmi. I build intelligent systems and teach the ideas behind them — connecting AI, data, and software with the people who use them.
Based in Nepal. Connected by curiosity.
Across the practice01Machine learning02Software & systems03Open education04Applied research
01 / Learn with me
A little clarity. A lot of possibility.
Open course notes for curious minds. Start with the fundamentals, follow the ideas, and put them into practice.
All teaching resources✳
Built for the classroom.
Open for everyone.
02 / Principles in practice
Clarity is a practice.
Three principles that guide how I teach, build, and explore.
Start with a clear mental model. Give the formalism something to attach to.
03 / Projects & research
A wider field of possibilities.
A project lab spanning agentic systems, Nepali language models, voice AI, and the data behind better decisions.
Stateful multi-agent systems
Conductor · Agent Runtime
An orchestration layer for specialist agents that plan, use tools, recover from failures, and hand work back to people.
PythonLangGraphPostgreSQL
Agentic AIConcept
Nepali LLM training & adaptation
Akshar · Nepali Language Foundry
A reproducible Nepali language-model program spanning corpus curation, Devanagari tokenization, continued pretraining, and instruction tuning.
PyTorchTransformersSentencePiece
LLMs & NLPResearch
Realtime Nepali–English conversation
Awaaz · Bilingual Voice Agents
A realtime voice-agent architecture for Nepali–English conversations, interruptions, tool calls, and human handoff.
LiveKit AgentsWebRTCPython
Voice AIConcept
04 / Articles & field notes
Thinking in the open.
Detailed guides and small observations on AI, mathematics, engineering, and the systems behind useful software.
019 min read029 min read039 min read
Agentic RAG: from retrieval to reliable reasoning
Build an evidence-driven retrieval loop with query planning, hybrid search, bounded tool use, grounded answers, and evaluation that exposes failures.
A language model from scratch: tokens, attention, and training
Derive next-token prediction and causal attention, then build a small decoder-only Transformer with a complete PyTorch training and sampling script.
Forecasting with mathematical intuition
Understand exponential smoothing, ARIMA, supervised forecasting, rolling backtests, and prediction intervals through equations and practical examples.
A common thread
Build to understand. Teach to understand better.
Engineering and teaching inform each other. Building reveals the questions worth asking. Explaining makes the answers clearer. This is a home for both.
A little more about me