AI & Generative AI App Development

What you will be able to do

• Write clean Python that calls web APIs, handles errors and keeps secrets out of code
• Explain how large language models work, what they are good at and where they fail
• Call an LLM API with streaming, retries, timeouts and cost tracking
• Design, test and iterate prompts that behave predictably on real inputs
• Get validated structured outputs from a model and handle the cases where validation fails
• Build semantic search with embeddings and a vector index
• Build a retrieval-augmented generation pipeline that cites its sources and refuses when it lacks evidence
• Evaluate an LLM application with a test set and catch regressions before release
• Defend against prompt injection and data leakage, and apply privacy, consent and cost limits
• Deploy an AI application with configuration, logging and monitoring suitable for real users

About this course

Syllabus

This syllabus is a draft awaiting instructor review. Topics, order and schedule may change before the first batch is announced.

1. 01Python foundations for AI applications+4 live sessions · 8 hours · The Python the rest of the course relies on: data structures, functions, files and JSON, calling web APIs and handling configuration and secrets safely.

   Learning objectives

   • Set up a Python project with a virtual environment, dependencies and Git
   • Use lists, dictionaries, functions, classes and type hints confidently
   • Read and write files and JSON, and call HTTP APIs with proper error handling
   • Keep API keys in environment variables and out of code and logs

   Live sessions

   1. 01Course setup and Python essentials2 hInstalling Python, VS Code and Git; virtual environments and dependency files; variables, strings, conditions, loops and the interactive interpreter; notebooks versus scripts.
   2. 02Functions, data structures and JSON2 hLists, dictionaries and comprehensions; functions, arguments and type hints; dataclasses; reading and writing text files and JSON; exceptions and when to catch them.
   3. 03Calling web APIs and handling secrets2 hHTTP requests and status codes with an HTTP client library, timeouts and retries with backoff, environment variables and .env files, never committing or logging keys, a first look at async code.
   4. 04Python review and doubt-clearing2 hRefactoring a messy script into small, typed functions together; common beginner mistakes; open questions.

   Practice

   Write a small command-line tool that loads a folder of text files, calls a public web API with retries and saves a JSON report, with the API key read from the environment.

   Assessment

   Quiz on Python data structures, exceptions and handling secrets, plus a short scripting assignment reviewed with feedback.

   You will be able to

   You can write clean, typed Python scripts that work with files and web APIs safely.

2. 02LLM APIs and prompt design+4 live sessions · 8 hours · How large language models behave, how to call them reliably from code and how to write prompts that produce predictable results.

   Learning objectives

   • Explain tokens, context windows, sampling and why models can produce confident but wrong answers
   • Call an LLM API with system instructions, conversation history, streaming and error handling
   • Track token usage and cost per request and set limits
   • Design prompts with clear instructions, examples and separated untrusted input, and iterate on them with evidence

   Live sessions

   1. 05How large language models work2 hTokens and context windows, next-token prediction in plain terms, temperature and sampling, what models do well, hallucination and knowledge cut-offs, choosing a model by task, latency and cost.
   2. 06Calling an LLM API from Python2 hMessages and roles, system instructions, conversation history, streaming responses, timeouts, retries and rate-limit errors, reading token usage, per-request and daily cost limits.
   3. 07Prompt design fundamentals2 hClear instructions and success criteria, few-shot examples, delimiting user input from instructions, asking for reasoning where it helps, prompt templates in code, versioning prompts like source code.
   4. 08Prompt lab and doubt-clearing2 hImproving weak prompts against a small set of test inputs, comparing outputs side by side, recognising when a prompt change cannot fix a task; open questions.

   Practice

   Build a summarisation and rewriting helper for the course project’s FAQ documents, with streaming output, a token budget and a prompt tested against ten sample inputs.

   Assessment

   Assignment: a prompt with its test inputs, before-and-after outputs and a short note explaining each change.

   You will be able to

   You can call an LLM API reliably from code and write prompts whose behaviour you can explain and test.

3. 03Structured outputs and AI-powered backends+3 live sessions · 6 hours · Get typed, validated data back from a model and wrap LLM features in a small, well-behaved web service.

   Learning objectives

   • Request structured outputs with JSON schemas and validate them with Pydantic
   • Handle validation failures, refusals and partial answers without crashing or guessing
   • Build extraction and classification pipelines that flag low-confidence results for human review
   • Expose an LLM feature through a FastAPI endpoint with input validation and consistent errors

   Live sessions

   1. 09Structured outputs with schemas and Pydantic2 hJSON mode and schema-constrained outputs, defining Pydantic models, validating and retrying on invalid output, enums and optional fields, never filling in missing facts with invented values.
   2. 10Extraction and classification pipelines2 hExtracting fields from documents, classifying text into fixed labels, batching and concurrency limits, confidence signals and routing uncertain cases to a person, logging without personal data.
   3. 11Building an AI backend with FastAPI and doubt-clearing2 hFastAPI routes and request models, calling the model from a service layer, streaming to the client, timeouts and graceful errors, rate limiting per user; open questions.

   Practice

   Add a policy-extraction endpoint to the course project that returns a validated record of key dates and obligations, with unit tests for valid, invalid and refused outputs.

   Assessment

   Project milestone 1: a FastAPI service with one structured-output feature and tests, reviewed with feedback.

   You will be able to

   You can turn unstructured text into validated, typed data and serve it from a reliable API.

4. 04Embeddings and vector search+3 live sessions · 6 hours · Represent text as embeddings, store them in a vector index and build semantic search that finds meaning, not just keywords.

   Learning objectives

   • Explain what embeddings capture and how similarity search works
   • Store and query embeddings in PostgreSQL with pgvector, with metadata filters
   • Choose an index type and measure the trade-off between speed and recall
   • Combine keyword and semantic search where each one helps

   Live sessions

   1. 12Embeddings and semantic similarity2 hWhat an embedding is, cosine similarity and distance, choosing an embeddings model, batching requests and caching embeddings, cost and re-embedding when the model changes.
   2. 13Vector search with PostgreSQL and pgvector2 hStoring vectors alongside metadata, nearest-neighbour queries, HNSW and IVFFlat indexes, filtering by source and access level, hybrid search with full-text ranking, other vector databases in brief.
   3. 14Search lab and doubt-clearing2 hBuilding and testing semantic search over the course documents, inspecting results that look wrong and why; open questions.

   Practice

   Embed the course project’s documents, store them with pgvector and build a search endpoint, then compare keyword, semantic and hybrid results on fifteen sample queries.

   Assessment

   Quiz on embeddings, similarity and vector indexes, plus a short search-comparison note.

   You will be able to

   You can build semantic search over a document collection and explain why it returns what it returns.

5. 05Retrieval-augmented generation+4 live sessions · 8 hours · Ground model answers in your own documents: ingestion, chunking, retrieval, reranking and answers that cite their sources.

   Learning objectives

   • Design an ingestion pipeline that cleans, chunks and indexes documents with useful metadata
   • Improve retrieval with chunking choices, query rewriting, hybrid search and reranking
   • Generate answers that cite sources and refuse when the context does not contain the answer
   • Respect document permissions so retrieval never leaks content a user should not see

   Live sessions

   1. 15RAG architecture and document ingestion2 hWhy retrieval beats stuffing everything into a prompt, loading PDFs and web pages, cleaning text, chunking strategies and overlap, metadata, keeping the index in sync when documents change.
   2. 16Improving retrieval quality2 hChunk size experiments, query rewriting, hybrid retrieval, reranking, how many passages to include, diagnosing retrieval failures separately from generation failures.
   3. 17Grounded answers, citations and permissions2 hPrompting for answers only from supplied context, citing passages, saying “I don’t know”, handling conflicting sources, filtering retrieval by user permissions, multi-turn questions.
   4. 18RAG review and doubt-clearing2 hReviewing each learner’s pipeline on hard questions; common failure patterns and fixes; open questions.

   Practice

   Build the course project’s question-answering endpoint with citations, a refusal path and permission filtering, and test it on questions the documents can and cannot answer.

   Assessment

   Project milestone 2: a working RAG assistant over the course documents, reviewed live with a set of unseen questions.

   You will be able to

   You can build a RAG application that answers from your documents, shows its sources and admits when it lacks evidence.

6. 06Evaluation, safety and responsible AI+3 live sessions · 6 hours · Measure answer quality instead of trusting it, and protect users, data and budgets from the ways LLM applications go wrong.

   Learning objectives

   • Build an evaluation set and automated checks for retrieval and answer quality
   • Use model-graded evaluation carefully, knowing its biases, and confirm it against human review
   • Recognise and mitigate prompt injection, data leakage and unsafe outputs
   • Apply privacy, consent, data-minimisation and human-oversight practices to an AI feature

   Live sessions

   1. 19Evaluating LLM applications2 hBuilding a test set from real questions, retrieval metrics such as recall and precision at k, answer checks for correctness and groundedness, model-graded evaluation and its limits, regression testing in CI.
   2. 20Security and safety for LLM applications2 hDirect and indirect prompt injection, untrusted content in retrieved documents, leaking system prompts and private data, output filtering, least privilege for anything the model can trigger, red-teaming your own application.
   3. 21Privacy, consent, cost and human oversight2 hCollecting the minimum data, consent and explaining what is sent to which provider, provider data-retention settings, redacting personal data in logs, India’s DPDP Act 2023 in outline, budgets and alerts, human approval for consequential outputs; doubt-clearing.

   Practice

   Write an evaluation set of thirty questions for the course project, run it against two configurations, and attempt a set of guided prompt-injection attacks on your own assistant and fix what succeeds.

   Assessment

   Assignment: an evaluation report with results for two configurations and a safety self-review with evidence of the defences you added.

   You will be able to

   You can measure whether an AI feature works, show the evidence and close the most common safety and privacy gaps.

7. 07Deployment and final project+3 live sessions · 6 hours · Package and deploy an AI application with monitoring and cost controls, then complete, present and defend the final project.

   Learning objectives

   • Containerise and deploy an AI service with environment-based configuration
   • Add logging, tracing, usage monitoring and alerts for errors, latency and spend
   • Apply the course’s evaluation, safety and privacy practices to a complete project
   • Present and defend design decisions and known limitations in a live review

   Live sessions

   1. 22Deploying and operating AI applications2 hA Dockerfile for a Python service, configuration and secrets in deployment, caching responses and embeddings, rate limits and budgets, structured logs and traces of model calls, monitoring quality after release, provider outages and fallbacks.
   2. 23Final project build lab and mid-point review2 hGuided build time on the final project with instructor support; a short review of each project’s architecture, evaluation plan and safety measures; agreeing what to fix before submission.
   3. 24Final project presentations and final assessment2 hLive project demonstrations with evaluation results and feedback; final assessment across the whole course (pass mark 60%); next steps for continued learning.

   Practice

   Deploy the course project with a usage log and budget limit, then complete the final project with regular commits, a working deployment and an up-to-date README.

   Assessment

   Final project submission and live review, and the final assessment (pass mark 60%).

   You will be able to

   You can independently deliver, evaluate, deploy and explain a responsible LLM application.

Teaching formats and fees

Every format teaches the full syllabus live, with the same projects, assessments and certificate rules. The fee is per student for the whole course.

Launch fee — GST treatment to be confirmed.

Fees are set in Indian rupees. Amounts in other currencies are estimates; the amount and currency you would pay are shown before payment once online enrolment opens.

Schedules and batches

Choose the schedule that fits your week. Both teach the same syllabus and the same 48 live hours; only the number of weeks differs.

• Weekday Intensive classes are one hour, so each 2 h syllabus session is taught across consecutive weekday classes.
• Weekend Evening classes run on Saturday and Sunday evenings; the exact time and time zone are set before a batch is confirmed.
• Public holidays are not class days unless the instructor confirms a class. Holidays, instructor leave and cancelled classes move the end date later so no live hours are lost.
• Plan about 5 hours a week of exercises, assignments and project work outside class.
• Group, Duo and One-to-One courses all follow one of these two schedules.

Final project

Certificate

Tools and skills

Where these skills apply

• Adding AI features to existing web and mobile applications
• Building internal assistants and search over company documents
• Document extraction, classification and summarisation pipelines
• Backend development for AI-enabled products
• Prototyping and evaluating AI ideas for a business before investing in them
• Freelance and small-team AI application development

We help you prepare for work; we do not guarantee jobs, placements or salaries. Career resources

Questions about AI & Generative AI App Development