LLMs and Generative AI for Beginners
What a large language model is, how to use one well and safely, a little code, and a capstone AI helper you build and evaluate, for anyone starting with AI
9 modules · 59 video lessons · about 10 hours
I'm interestedAbout the course
A first course on large language models and generative AI in nine modules, with a clear example in every lesson and no coding background assumed. It starts from the AI already around you (the spam filter, the phone keyboard, the maps app), then opens the model up in pictures: tokens, next-token prediction, temperature, the context window, how a model is trained, embeddings and attention. Then prompts that work, the real tasks of work and study (summaries, letters, translation between English, Hindi and Odia, study plans, spreadsheets), trusting the answer (hallucinations, verification, privacy, prompt injection, honest use), giving the model your own documents (retrieval and RAG), a little Python in Google Colab to call a model from code, and how AI apps are built with tools, agents and evaluation.
You make things as you go: a prompt library, a cited document answer, the Notice Explainer (a government notice in plain English, Hindi and Odia as JSON) and a twenty-question test set. The course ends with a capstone AI helper of your own (the Exam Prep Buddy, or a shop assistant or a notice desk) on a no-code or a code route, evaluated before and after, checked for safety and scored against a 100-point rubric, then a final test.
- Who it is for
- College students of any branch, freshers and working people who want to start with AI. No coding background is needed until module 7, which teaches just enough Python in a browser notebook; the capstone has a no-code route too. A phone is enough for every lesson; a laptop helps for modules 7 to 9.
- Format
- Recorded video lessons of about ten minutes with slides, worked examples and narration in English, Hindi and Odia. Slides, code and quizzes are in English. Not a live class.
- Length
- 9 modules · 59 video lessons · about 10 hours. About an hour a day: a lesson of about ten minutes, its try-it task and the lesson quiz.
- Free lessons
- All 6 lessons of module 1. Their notes and quizzes are open now; the videos are being prepared.
- Price
- ₹1,999, paid once. Launch price ₹1,199 for the first 100 students. Enrolment opens soon; no payment is taken now.
What you will be able to do
- Explain what a large language model is and how it works, with an everyday picture for each idea: a token, next-token prediction, temperature, the context window, training and the cutoff, embeddings, attention.
- Write a prompt with the task, the context, the format, a role and an example, ask for a fixed output shape, and improve a reply with follow-ups.
- Do the real tasks of work and study with an assistant (summaries, extraction, letters, translation, a study plan, a spreadsheet formula) and check each result before using it.
- Spot a hallucination, run a four-step verification, keep private data out of a prompt, recognise prompt injection and an AI scam, and say honestly where AI helped.
- Give a model your own documents, judge a cited answer, and explain retrieval-augmented generation.
- Call a chat model from Python with the key kept secret, loop over twenty items, ask for JSON and check it, and estimate the cost in rupees.
- Explain how an AI app is put together, what tools and agents are, how to choose a model, and how to evaluate a feature with a test set and a rubric.
- Finish a capstone AI helper scored against a 100-point rubric, with its before-and-after scores recorded, and take the final test.
Curriculum: 9 modules, 60 lessons
59 video lessons, about 10 hours in all. Open a module to see its lessons.
Module 01 AI around you: what generative AI is 6 lessons
No coding background needed. What artificial intelligence is and is not, told through the spam filter, the phone keyboard and the maps app; what generative means; your first conversations with a chat assistant; how a language model learns from text; what it does well and badly; and the words everyone uses, decoded.
-
What AI is, and what it is not Free 9 min
Tell a program that follows written rules from one that learned patterns from examples (the spam filter, the maps app, the phone keyboard), and sort six everyday apps into the two kinds.
-
Generative AI: machines that make new things Free 9 min
Tell recognising from generating (reading a number plate versus writing a letter), name the four things generative AI makes (text, images, audio, code), and say what a large language model is in one sentence.
-
Your first conversation with a chat assistant Free 10 min
Open a free chat assistant on a phone or a browser, ask a question, follow it up, correct it, and start a new chat when the topic changes, with the three habits that make answers better.
-
How a language model learns: patterns from text Free 10 min
Explain with the phone keyboard's next-word suggestion how a model learns from a huge amount of text to predict the next piece, and why it writes fluent sentences without looking anything up.
-
What LLMs do well and badly Free 10 min
Sort ten everyday tasks into ones an assistant does well (drafting, summarising, explaining, translating) and ones it gets wrong (exact facts, recent events, arithmetic, your private data), and say why for each.
-
The words everyone uses: model, parameters, training, inference, GPT Free 9 min
Decode the words in every AI headline (model, parameters, training, inference, prompt, GPT, open and closed models) with one everyday picture each, and read a model announcement without being lost.
Module 02 Inside a language model: tokens, prediction and context 7 lessons
The picture version of how a model works, with a clear example in every lesson: tokens (how a model reads), next-token prediction, temperature and randomness, the context window (the model's short-term memory), the three stages of training and the knowledge cutoff, meaning as numbers (embeddings) and attention.
-
Tokens: how a model reads text 10 min
Split a sentence into tokens the way a model does, estimate the tokens in a paragraph (about four English characters each, more in Odia and Hindi), and say why the count matters for cost and limits.
-
Predicting the next token, one piece at a time 10 min
Explain that a model picks the next token by probability (the capital of a state is ...), read a bar chart of candidate tokens, and say why an answer appears word by word on the screen.
-
Temperature: why the same question gets different answers 9 min
Explain sampling with a weighted lucky draw, predict what a low and a high temperature do to a reply, and choose the setting for a legal summary versus a story.
-
The context window: the model's short-term memory 10 min
Explain that the whole conversation is re-read on every turn, estimate when a long chat will run out of room, and say what to do when the model forgets what you said earlier.
-
How a model is trained: three stages and a cutoff 10 min
Describe pre-training, instruction tuning and learning from human feedback with the apprentice analogy, say what the knowledge cutoff is, and explain why an assistant is polite and follows instructions.
-
Meaning as numbers: embeddings 10 min
Explain how a model turns a word or a sentence into a list of numbers so that similar meanings land close together (fees near tuition charges), and read a two-dimensional map of words.
-
Attention and transformers, the picture version 10 min
Explain attention as the model deciding which earlier words matter for the next one (what 'it' refers to), say what a transformer is in one sentence, and why bigger models need more computing.
Module 03 Talking to a model: prompts that work 7 lessons
The prompt is the only control a user has. Task, context and format; roles and audiences; examples that show the answer you want; asking for lists, tables, JSON and word limits; step-by-step reasoning; follow-ups and corrections; and a prompt library of five templates you keep.
-
The anatomy of a prompt: task, context, format 10 min
Rewrite a one-line request (write a leave letter) into a prompt with the task, the context and the format, and predict which of two prompts gives the more useful reply.
-
Give the model a role and an audience 9 min
Set a role (a careful tutor, a bank clerk) and an audience (a Class 8 student, a shop owner), and predict how the same explanation changes for three audiences.
-
Show, do not just tell: examples in a prompt 10 min
Add two examples to a prompt so that a messy list of addresses comes back in one consistent format, and say when one example is enough and when it misleads.
-
Ask for a format: lists, tables, JSON and word limits 10 min
Ask for a numbered list, a table, a 50-word limit or JSON with named fields, read the JSON back, and say why a program needs that fixed shape.
-
Think step by step: getting better reasoning 10 min
Ask the model to work through a problem in steps before answering (a bill with a discount and tax), check each step, and spot the step where a wrong answer went wrong.
-
Follow-ups, corrections and starting over 9 min
Improve a reply with follow-ups (shorter, simpler words, add an example), correct a wrong assumption, and decide when to start a fresh chat instead.
-
Mini project: your prompt library 11 min
Build five reusable prompt templates (an email, a summary, a study plan, a translation, explain a topic) with blanks to fill, test each once, and keep them in a notes file.
Module 04 GenAI for work and study: real tasks, done well 6 lessons
The tasks students and working people actually bring to an assistant, each done with a clear example and checked: summarising and extracting, writing and rewriting, translating between English, Hindi and Odia, learning with an assistant, spreadsheets and quick analysis, and images, audio and code.
-
Summarising and extracting: long text to the points you need 10 min
Summarise a two-page notice to five points, extract the dates, names and amounts into a table, and check the table against the text.
-
Writing and rewriting: emails, applications and notices 10 min
Draft a job application email from bullet points, change its tone (formal, friendly, shorter), fix grammar without changing meaning, and keep your own facts in it.
-
Translating and simplifying: English, Hindi and Odia 10 min
Translate a paragraph between English, Hindi and Odia, simplify a government circular into plain words, and check a translation by translating it back.
-
Learning with an assistant: explanations, quizzes and study plans 10 min
Use an assistant as a tutor: ask for an explanation at your level, a five-question quiz on it, and a two-week study plan, and avoid the trap of letting it do the work for you.
-
Spreadsheets and quick analysis 10 min
Ask for a spreadsheet formula and explain what it does, paste a small table of marks and ask for the average and the top three, and check the numbers by hand.
-
Images, audio and code: the other generative tools 10 min
Write an image prompt with subject, style and setting, use speech-to-text and text-to-speech for a short note, and ask for a small program as a non-programmer, knowing what to expect from each.
Module 05 Trusting the answer: hallucinations, privacy and honest use 6 lessons
Why a fluent answer can be wrong and what to do about it: hallucinations, a verification routine, bias, what never goes into a prompt, prompt injection and AI scams, and using AI honestly at school and at work.
-
Hallucinations: confident and wrong 10 min
Explain from next-token prediction why a model invents a fact, a date or a citation, spot three hallucinations in a sample answer, and say which kinds of question are most at risk.
-
Verify before you use: sources, cross-checks and uncertainty 10 min
Run a four-step check on an answer (ask for sources, open one, cross-check with a second source, ask the model how sure it is), and decide when an answer may be used without checking.
-
Bias and fairness: where the training data shows 9 min
Explain how training data puts stereotypes into answers, spot a biased reply in three examples, and rewrite a prompt so the answer treats people fairly.
-
Privacy: what never goes into a prompt 10 min
List what never goes into a prompt (Aadhaar and PAN numbers, passwords, bank details, patient records, company secrets), say what a service may keep from a chat, and find the settings that control it.
-
Prompt injection and AI scams 10 min
Explain how hidden text in a document can make an assistant follow a stranger's instructions, and recognise a voice-clone or deepfake scam with three checks before acting.
-
Using AI honestly: school, work and credit 9 min
Decide what to disclose when AI helped with an assignment, a report or a design, tell help from plagiarism, and follow a five-point honesty checklist.
Module 06 Giving the model your own knowledge: documents, search and RAG 6 lessons
A model knows nothing about your file. Three ways to give it knowledge, chatting with a document, retrieval (finding the right pages first), meaning search, chunks and citations, and web search inside an assistant, with the librarian analogy throughout.
-
Why the model does not know your file, and three ways to fix it 9 min
Tell what is in the training data from what is in the context, and choose between pasting, uploading and connecting a source for a question about your own syllabus.
-
Chat with a document: upload and ask 10 min
Upload a PDF (a syllabus, a notice), ask questions that are answered on its pages, ask for the page each answer came from, and check one against the page.
-
Retrieval: finding the right pages first 10 min
Explain retrieval-augmented generation with the librarian (a question, a search, the pages found, the answer written from them), and trace a question through the four steps.
-
Meaning search: beyond keywords 10 min
Explain why a search by meaning finds 'tuition charges' for 'fees' when a keyword search does not, using the map of embeddings from module 2, and say what a vector store does in plain words.
-
Chunks and citations: why answers go wrong and how to fix them 10 min
Explain why a chunk that is too big or too small spoils an answer, ask for a citation with every claim, and judge a document answer on three points (found, faithful, complete).
-
Web search inside an assistant: live answers and their limits 9 min
Use an assistant's web search for a recent fact, open the linked source, and say when a search-grounded answer still needs checking.
Module 07 A little code: your first AI program in Python 8 lessons
Why a few lines of code change what you can do with a model, then just enough Python in Google Colab (nothing to install): values, variables and strings, your first API call with the key kept secret, messages and roles, the same prompt for twenty items, JSON you can check, tokens, cost and errors, ending with the Notice Explainer mini app.
-
Why a little code changes everything 9 min
Say what code adds to a chat window (the same task a hundred times, your own data, a program others can use), open Google Colab in a browser, and run one line.
-
Python in ten minutes: values, variables and strings 11 min
Store text and numbers in variables, join text with an f-string, and predict what three short programs print, the minimum needed to talk to a model.
-
Your first API call: a key, a request, a reply 11 min
Explain what an API is with the bank counter, keep the API key in a Colab secret and never in the code, send one prompt to a chat model and print the reply.
-
Messages and roles: system, user, assistant 10 min
Build the list of messages a chat model reads (a system message, user turns, assistant turns), set the assistant's behaviour in the system message, and predict how the reply changes.
-
The same prompt for twenty items: lists and loops 10 min
Loop over a list of product names and ask the model for a one-line description of each, and predict how many calls a loop makes.
-
Ask for JSON and check it 10 min
Ask the model for a reply in JSON with named fields, turn it into a Python dictionary, and check that every field is present before using it.
-
Tokens, cost and errors in code 10 min
Count the tokens of a request and a reply, estimate the cost of a thousand calls in rupees, and handle a rate limit or a network error by waiting and retrying.
-
Mini app: the Notice Explainer 12 min
Build a program that takes a government notice and returns a plain-language summary in English, Hindi and Odia as JSON with the dates and amounts extracted, and test it on two made-up notices.
Module 08 How AI apps are built: tools, agents and choosing a model 6 lessons
What is inside an AI app and where the cost goes, tools and function calling (the model asks, your program acts), agents as a loop of think, act and check, choosing a model (small or big, local or cloud, when not to fine-tune), evaluating an AI feature with a test set and a rubric, and what to learn next.
-
The anatomy of an AI app 10 min
Name the four parts of an AI app (a system prompt, a model, your data, a screen), say what each costs, and trace one user question through them.
-
Tools and function calling: the model asks, your program acts 10 min
Explain how a model asks for a calculator, a weather lookup or a database query, and why your program, not the model, runs it; trace a question that needs a tool.
-
Agents: a loop of think, act and check 10 min
Describe an agent as a loop (plan, use a tool, look at the result, repeat), follow a travel-planning agent through three rounds, and name the two places it goes wrong.
-
Choosing a model: small or big, local or cloud 10 min
Compare models by size, cost, speed and where they run, say what open weights means, and explain why a better prompt or retrieval usually beats fine-tuning for a beginner.
-
Evaluating an AI feature: a test set, a rubric and a score 11 min
Write a test set of twenty questions with expected answers, score replies on a three-point rubric, and compare two prompts by their scores instead of by feel.
-
What to learn next: roles, paths and a 90-day plan 9 min
Tell the roles that use these skills from the ones that build the systems, see how the track's GenAI & LLM Engineering exam continues from here, and write a 90-day plan.
Module 09 Capstone: build and evaluate your own AI helper 8 lessons
Your own AI helper from brief to demo: the Exam Prep Buddy (or two alternatives), a no-code route and a code route, planned with a test set, built, evaluated before and after, checked for safety, scored against a 100-point rubric and presented, then the course's final test.
-
The capstone brief: an AI helper for a real need 12 min
Read the capstone brief: the Exam Prep Buddy (or the Shop Assistant or the Notice Desk), the must-haves, the two routes (no-code and code), the milestones and the rubric.
-
Plan: the users, the twenty questions and the knowledge 11 min
Write who the helper is for, the twenty questions it must answer with their expected answers, the knowledge file it needs and the three rules it must follow, before building anything.
-
Milestone 1: the system prompt and the knowledge file 11 min
Write the helper's system prompt (role, rules, format, refusals), prepare and chunk the knowledge file, and test both on five of the twenty questions.
-
Milestone 2: build it 12 min
Build the helper on your route: a custom assistant with the prompt and the uploaded knowledge (no-code), or a Colab notebook that retrieves from the file and chats (code), and run all twenty questions.
-
Milestone 3: evaluate and improve 11 min
Score the twenty answers with the rubric from module 8, fix the prompt or the knowledge where it lost points, score again, and record the before and after.
-
Safety review and the 100-point rubric 11 min
Check the helper for privacy, injection and honesty (what it refuses, what it discloses), then score the project against the 100-point rubric and fix the gaps.
-
Presenting your project 10 min
Demonstrate the helper in two minutes, write the one-page report (what it does, the scores, what it must not be used for), and write the project line for your resume.
-
Course final test Final test
Ten mixed questions in 15 minutes across the course, no negative marking, to see which modules to revise before the track's full mocks.
Capstone project
You build an AI helper of your own that answers from a knowledge file you prepare, explains its answers, and refuses questions outside its knowledge instead of guessing. There is a no-code route and a code route, scored by the same 100-point rubric.
Projects (choose one, or propose your own)
- Exam Prep Buddy: a helper for one exam or one subject you know well
- Shop Assistant: a family shop's products, prices, timings and return rules
- Notice Desk: an office's or a college's notices, explained in plain words
Every project has
- A written plan: who it is for, twenty questions with expected answers, the knowledge file and three rules
- A system prompt you wrote yourself, with a role, the rules, the answer format and what it refuses
- A knowledge file of your own making, split into chunks that make sense
- Answers from the knowledge, cited when asked, and refusals outside it
- Evaluated twice with a three-point rubric, before and after one round of improvement
- A safety check: no private data, one injection attempt recorded, a line saying it is an AI helper and may be wrong
- A one-page report and a two-minute demo
Free practice
Questions
Are these live classes?
No. Every lesson is a recorded video of about ten minutes with slides, worked examples and narration, so you study at your own pace. The lessons are written and checked by our team, and the narration is computer generated. This is different from our internships, which are taught live.
What is free?
The whole first module of each course: its notes and quizzes are open now, and its videos are being prepared. The rest comes with the course.
Can I buy a course today?
Not yet. Enrolment opens soon. Leave your details in the form on the course page and we will write to you when it opens.
What is the launch price?
The lower price shown on each course page, for the first 100 students. After that the list price applies. You pay once; there is no subscription.
Which languages are the lessons in?
The narration is in English, Hindi and Odia. Slides, code and quizzes are in English.
Do I need to install anything?
Not to start. The first lessons run in a browser. The Python Full Stack course needs a laptop or a desktop from its third lesson on, and tells you what to install when you need it.
How is my work checked?
Lessons come with short quizzes. The capstone project is scored by you against a published 100-point rubric, and each course ends with a final test.
