AI & Generative AI

The mental model everything else rests on. No maths, no code.

• Rules versus learned behaviour
  Why ordinary software follows rules a person wrote and a model works the pattern out from examples — and why that one difference explains almost every strength and every failure that follows.
• Training, inference and why they cost differently
  What actually happens when a model is trained versus used. Why using AI is cheap and building it is not, and which of the two most jobs involve.
• Where AI genuinely earns money in Pakistan today
  Support answering, document reading, drafting, sorting enquiries. Unglamorous, high volume, human-checked — and hiring.
• The four ways it fails hands on
  Confident invention, absorbed bias, silent drift, and looking right to someone who cannot check. Naming them is the skill.

Getting reliable output from ChatGPT, Gemini and Claude.

• The anatomy of a prompt that works hands on
  Role, context, constraints, format and examples — and why 'answer the questions a colleague would have asked' beats every magic phrase.
• Showing beats describing hands on
  Few-shot prompting: why two examples of the output you want outperform a paragraph describing it.
• Controlling the format so output is usable hands on
  Getting JSON, tables and fixed structures out reliably, so the result can feed something else instead of being retyped.
• Chaining and decomposition hands on
  Breaking a task the model fails at into steps it succeeds at.
• Testing prompts instead of fiddling with them
  Building a small set of cases and measuring, so 'it seems better' becomes something you can defend to a client.

The single most requested business use, and the one most people get wrong.

• Why a general model cannot answer about your business
  It never saw your price list, your policy or your handbook — and will cheerfully invent all three.
• Grounding: retrieve, then answer
  The pattern behind every document assistant: find the relevant text first, hand it to the model, and forbid it from going beyond that.
• Chunking and search that actually finds the right passage hands on
  Keyword versus embedding retrieval, and the honest guidance on when a simple approach is genuinely better.
• Refusing rather than guessing hands on
  Designing the answer 'I do not have that' — the feature that decides whether a business can trust the system.

Connecting models to real workflows with n8n and APIs.

• Triggers, actions and the shape of an automation hands on
  Reading the workflow before building it: what starts it, what it touches, what it must never do twice.
• Calling AI from an automation hands on
  API keys, request and response, cost per call, and handling the call that fails.
• Idempotency and the double-send problem hands on
  Why an automation that runs twice must not email a customer twice — the failure that embarrasses clients most.
• Human in the loop
  Where to put an approval step, and how to decide which decisions a business should never fully automate.

Generative media, and the professional limits on using it.

• Image generation and editing for real briefs hands on
  Product shots, social creative and mockups — where it saves hours and where it still cannot be trusted.
• Text to speech, including Urdu hands on
  Producing narration in Urdu and English, and judging when a synthetic voice is good enough to publish.
• Disclosure and client consent
  When AI-generated media must be declared, and why an undisclosed synthetic likeness is a problem you cannot apologise your way out of.

Putting it in front of real users without a nasty surprise.

• What a call actually costs hands on
  Tokens, per-call pricing, caching, and estimating a monthly bill before a client asks.
• Rate limits and graceful degradation hands on
  Designing what the user sees when the quota is gone — the difference between a broken product and a busy one.
• Data you must never send to a third-party model
  Customer records, CNIC numbers, credentials, anything under a confidentiality clause. The practical rules.
• Guardrails and output checking hands on
  Why an instruction in a prompt is a request and a filter in code is a rule, and how to test that the rule holds.

Pakistan's National Vocational & Technical Training Commission publishes a course-contents document for the trades it funds. This programme is mapped to AI and Robotics — 12 weeks, 4 hours a day, 240 contact hours.

Published by NAVTTC and reproduced here so you can check our syllabus against it. We teach it to those hours: 240 contact hours over 12 weeks, the same as the standard. Read the NAVTTC document.

Before you start

• Comfortable using a computer and the internet
  No programming background is required for this programme.
• Free: AI Fundamentals
  The free mini-course covers the first module's ground; taking it first is recommended, not required.

Where it leads

• AI Automation & n8n
  Goes deeper into building automations for clients.
• Machine Learning
  For those who want to build models rather than apply them. Requires real mathematics.
• Python Programming
  The route to building AI products rather than configuring them.

Roles this prepares you for

• AI Automation Specialist
  Builds and maintains AI-driven workflows for businesses.
• AI Solutions Associate
  Applies generative AI inside an existing team — marketing, support or operations.
• Prompt & Content Systems Specialist
  Designs the repeatable prompt systems a content or support team runs on.

Interested in this programme?

Try the subject free first — there is no reason to pay before you know you like it.