AI FOUNDATIONS
A practical map of AI: from machine learning to agents
A guide for learners and businesses in Nepal and Kathmandu: how AI, ML, deep learning, NLP, vision and automation fit together, with links to my public projects.
AI agents and automation: connect a task to an outcome
Artificial intelligence is a broad field, and business automation is one application of it. An agent combines a model with tools and workflow logic to help with a task. An automation defines how information moves from a trigger to an outcome. They can work together, but they are not interchangeable terms.
For a team in Kathmandu or elsewhere in Nepal, start with the workflow you want to improve. A customer enquiry, a follow-up and a completed order are different steps. My work through AI Automation Nepal connects agents, n8n and software around those practical processes.
Generative AI and language models: produce and check
Generative models produce outputs such as text, images or audio from an input. Language models are useful for drafting, interpreting requests and working with unstructured text. A fluent response is not necessarily a correct one, so the application needs relevant context, checks and a way to handle uncertainty.
An integration can ask a model to return structured information and validate the result before using it. For consequential actions, the surrounding application should control permission and confirmation. Generating an answer and changing a business record are different responsibilities.
Machine learning: learn a pattern from examples
Machine learning uses data to fit a model for a task, such as estimating a value or assigning a category. A model should be evaluated on data that was not used to fit it. The right metric depends on the problem; a single attractive accuracy figure does not explain the cost of different mistakes.
My public laptop price prediction repository is one example of a regression project. It describes a random-forest model using laptop specifications and Nepal-focused data. It is a project to inspect and learn from, not a promise of current market prices or guaranteed predictions.
Deep learning: neural networks for complex inputs
Deep learning uses layered neural networks to learn representations from data. It is used in many image, language and audio tasks. Choosing an architecture is only part of the work: preprocessing, training data and a representative evaluation set strongly affect whether a model is useful.
My public projects include CNN-based classification and a MobileNetV2-based Nepali food classifier. The food project covers nine categories, including momo, dalbhat and selroti. These are visible examples of model-building work; their repository documentation provides the project-specific details.
NLP and Nepali language: work with text carefully
Natural language processing includes tasks such as classifying text, extracting named entities and representing words or passages for comparison. Tokenisation turns text into units a model can work with. For Nepali and mixed-language text, evaluate actual examples from the intended audience rather than assuming English results will transfer unchanged.
My public NLP work includes sentiment analysis, named entity recognition and a Nepali tokenizer repository. A DistilBERT-based sentiment model is published on Hugging Face. Read the model and project documentation to understand the task and inputs; language coverage and quality need their own evaluation.
Computer vision and OCR: understand images
Computer vision includes assigning a category to an image, locating objects and identifying regions. Optical character recognition concerns text in images; a character classifier is one component and does not by itself establish complete document-reading capability. Image quality and unfamiliar conditions can change performance.
My public work includes Devanagari character recognition, Nepal vehicle plate detection and food-image classification. Model repositories and demo Spaces provide concrete places to inspect that work. A demo may sleep or become unavailable, and a public upload alone is not independent validation of a model’s accuracy.
RAG: connect a question to relevant information
Retrieval-augmented generation combines information retrieval with a generative model. The system finds relevant material and supplies it as context for an answer. This can make a knowledge assistant more useful, but relevance, permissions and source quality still need attention.
Start with the documents and questions the intended users actually have. Test whether retrieval finds the correct information, whether the answer is supported and what happens when the information is missing. RAG is an engineering pattern, not a guarantee against incorrect answers. This section is an explainer, not a claim that every portfolio project uses RAG.
Speech and voice AI: evaluate the whole interaction
Speech recognition converts spoken audio into text, while speech synthesis produces audio from text. A voice interface adds further concerns: turn-taking, response time, noise and what happens when someone is misunderstood. Nepali, accents and code-switching should be tested with suitable audio.
Choose evaluation material that reflects the planned use, and get appropriate permission for recordings. Word-level errors and the success of the overall task are different measures. This is an area worth studying alongside language and agent systems, without assuming that a model demonstration establishes readiness for every caller.
AI engineering: the system around the model
An application needs more than a model file. Interfaces, APIs, access control, deployment and operational checks determine how people can use it. Measure failures as well as successful requests and document which outputs require human review. The appropriate safeguards depend on the use case.
For Nepal-focused projects, consider language, the devices people use and the business process alongside model choices. My public code and models show parts of that journey; my current company work brings AI agents, automation and software together. Explore the project links and start from the actual problem you need to solve.