25-Day Programme · 30 Minutes Per Day
WEEK 4 OF 5 · SESSIONS 16–20 · WORK AT YOUR OWN PACE
Domain 2 · Natural LanguageDAY 16
Text Analysis: Sentiment, Entity Detection, Key Phrases, Summarisation
Why This Matters
Welcome to Week 4, both of you. Last week you deployed a model and talked to it — this week you make AI do specific, sellable jobs. Today is about text. Every business swims in text it cannot read fast enough: customer reviews, WhatsApp messages, support emails, survey replies. A person who can point AI at a pile of text and pull out how customers feel, what they mention, and a one-line summary is immediately useful to any company. That is a real service you could sell in Lagos or Abuja tomorrow. Learn these four moves well.
30-Minute Module
0:00 – 6:00
What “text analysis” means and why it exists
- Natural language processing (NLP) is the branch of AI that works with human language — the words people actually write and speak. “Natural language” just means ordinary human language, as opposed to code or numbers.
- Text analysis is NLP applied to a block of writing to pull out structured facts from it. You hand it messy, free-form text; it hands back tidy, usable information.
- On the exam and in Foundry, four text-analysis capabilities come up again and again: sentiment analysis, entity detection, key phrase extraction, and summarisation. Learn these four names cold — they are Domain 2 bread and butter.
- The point is scale. A human can read ten reviews; AI can read ten thousand in the time it takes you to make tea. That leverage is what an employer pays for.
NLP = AI for human language. Text analysis = pulling structured facts (feelings, names, topics, summaries) out of unstructured writing. Say both definitions aloud before moving on.
6:00 – 16:00
The four capabilities, one at a time
- Sentiment analysis judges the emotional tone of text — positive, negative, or neutral — usually with a confidence score. “The delivery was fast and the phone works perfectly” is positive; “It arrived cracked and nobody replied” is negative. This is how a business reads the mood of a thousand reviews at once.
- Entity detection (also called named entity recognition, NER) finds and labels the real-world things named in text — people, places, organisations, dates, quantities. From “I bought a Tecno phone in Ikeja on Monday” it pulls Tecno (product/organisation), Ikeja (location), Monday (date).
- Key phrase extraction pulls out the main talking points — the handful of phrases that tell you what a text is about without reading it all. From a long complaint it might return “late delivery”, “damaged screen”, “no response”.
- Summarisation shortens a long text into a few sentences (this is extractive or abstractive — extractive picks key sentences out; abstractive writes fresh, shorter ones). Turning a 500-word email thread into three lines a busy manager can read is a genuinely valuable skill.
A memory hook: sentiment = how they feel; entities = what things they named; key phrases = what it is about; summarisation = the short version. Four questions, four tools.
16:00 – 25:00
Where this lives in Foundry — and language models can do it too
- There are two honest routes to these capabilities, and the exam expects you to know both exist. First, Azure offers ready-made Azure AI Language services — pre-built models where you send text and get sentiment or entities back with no training on your part. These are “prebuilt” capabilities: the work is already done, you just call them.
- Second, a general large language model (the GPT-family model you deployed on Day 13) can also do all four jobs if you ask it well — this is Day 12’s prompt engineering in action. A prompt like “Classify the sentiment of this review as positive, negative or neutral, then list the key phrases” gets you there.
- Knowing when to reach for a prebuilt service versus a general model is real judgement: prebuilt services are consistent and simple for a fixed job; a general model is flexible and handles fuzzy, combined requests. This is the same capability-versus-cost thinking from Day 9.
- Everything you run still happens inside your resource group —
rg-yourname-ai901, East US — against your subscription. Nothing new to provision today; you already have a deployment from Day 13.
Two roads to the same place: a prebuilt Azure AI Language capability, or a well-prompted general model. The exam wants you to know both exist and roughly when each fits.
25:00 – 30:00
Fix it in your own words
- Out loud, name the four text-analysis capabilities and the one-word question each answers: sentiment (feel), entities (things named), key phrases (about), summarisation (short version).
- In your AI-901 Notes, write one real business in your area for each capability — e.g. “sentiment: a phone shop reading its Jumia reviews”. Grounding it in reality is how it sticks.
- Tomorrow you build a small text-analysis app that actually does this in Python. Today, just make sure the four names and their jobs are solid in your head.
If any of the four still feels blurry, that is normal — report it to me today rather than sitting on it. A five-minute question now saves a stuck hour tomorrow.
Key Terms
Sentiment Analysis
Judging the emotional tone of text — positive, negative, or neutral — usually with a confidence score. Reads the mood of many messages at once.
Entity Detection (NER)
Finding and labelling real-world things named in text — people, places, organisations, dates, quantities. Also called named entity recognition.
Key Phrase Extraction
Pulling out the main talking points of a text — the few phrases that tell you what it is about without reading all of it.
Summarisation
Shortening long text into a few sentences. Extractive picks key sentences out; abstractive writes fresh, shorter ones.
Today’s Assignment
Analyse three real texts by hand and map each to the right capability.
- Find three short pieces of real text from your own life — a product review, a WhatsApp message from a business, a news paragraph. Paste each into your answer.
- For each one, do all four jobs yourself, on paper: (1) say its sentiment and why; (2) list the entities you find and label each (person/place/organisation/date); (3) write two or three key phrases; (4) write a one-sentence summary.
- Then write two or three sentences on this judgement question: for a phone shop that wants sentiment on 5,000 reviews every week, would you use a prebuilt Azure AI Language capability or a general GPT-family model, and why?
Submit assignment 16 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. A shop wants to know whether each of its 2,000 online reviews is positive or negative. Which capability fits?
- A. Entity detection
- B. Sentiment analysis
- C. Summarisation
- D. Key phrase extraction
B. Sentiment analysis — Positive/negative/neutral is exactly what sentiment analysis returns. Entities find named things, key phrases find topics, and summarisation shortens — none of those directly answer “how does the customer feel?”
2. From “MTN opened a new office in Abuja in March”, a model returns MTN=Organisation, Abuja=Location, March=Date. What capability is this?
- A. Key phrase extraction
- B. Sentiment analysis
- C. Entity detection (named entity recognition)
- D. Image generation
C. Entity detection — Finding and labelling real-world things (organisations, locations, dates) named in text is entity detection, also called named entity recognition. Key phrases would be topics like “new office”, not labelled entities.
3. Which statement is true about getting these capabilities in Azure?
- A. Only a prebuilt Azure AI Language service can do them; a general model never can
- B. Only a general GPT model can do them; prebuilt services do not exist
- C. Either a prebuilt Azure AI Language capability or a well-prompted general model can do them
- D. They can only be done by training your own model from scratch
C. Either route works — Azure offers prebuilt Language capabilities where the model is ready-made, and a general GPT-family model can do the same jobs with good prompting. Choosing between them is the capability-versus-cost judgement from Day 9.
Exam Objectives Covered
Natural language processingSentiment analysisEntity / NERKey phrase extractionSummarisationPrebuilt vs general model
Domain 2 · Build (Python)DAY 17
Build a Lightweight Text-Analysis Application
Why This Matters
Yesterday you learned the four text-analysis moves. Today you make them run in code — a small program that takes real text and returns real analysis. This is the day theory becomes a thing you built. Nothing impresses an employer, or a client, like being able to say “I wrote a program that reads reviews and tells you the sentiment” — and then show it working. Type every line yourself. The four-step pattern from Day 5 is your map today: Import → Connect → Send → Use.
30-Minute Module
0:00 – 6:00
Recall the four-step AI pattern
- Every program that talks to an AI service follows the same shape you met on Day 5: Import the library, Connect using your endpoint and key, Send your text and prompt, Use the answer that comes back. Four steps, always.
- Open
colab.research.google.comand start a new notebook — browser only, nothing to install, exactly like before. - Today you will point that pattern at your Day-13 deployment and ask it to do the four text jobs from yesterday. You already have the endpoint and key you saved privately on Day 13 — have them to hand, but do not paste them anywhere public.
- A word on keys before we touch code: your key authorises real spending on your subscription. In real projects you keep it out of the code itself. Today, for learning, you will paste it into a variable in your own private notebook — never into a shared or public one, and never into an email or screenshot.
Import → Connect → Send → Use. Write those four words at the top of your notebook as a comment before you write any code — they are the skeleton of everything today.
6:00 – 16:00
Import and Connect — type this yourself
- Install and import the client library. In a Colab cell, type (do not paste):
!pip install openai - Then in the next cell:
from openai import AzureOpenAI - Now Connect. Store your details in variables — put your own endpoint and key here:
endpoint = "https://YOUR-endpoint-here"key = "YOUR-key-here"deployment = "my-chat-model" - Create the client that will carry your messages:
client = AzureOpenAI(azure_endpoint=endpoint, api_key=key, api_version="2024-06-01") - If a name looks unfamiliar —
AzureOpenAI,api_version— that is fine. A client is just an object that knows how to reach your model; api_version tells the service which set of rules to use. You are wiring a phone line to your deployment.
If a cell errors, read the last line of the red message first — it usually names the problem in plain English (a typo, a wrong key). Being stuck is normal; report a blocker rather than staring at it silently.
16:00 – 26:00
Send and Use — make it analyse text
- Put the text you want to analyse in a variable:
review = "The phone arrived late and the screen was cracked, but support fixed it quickly." - Now Send. Give a system prompt (Day 12) that sets the job, and a user prompt that carries the text:
resp = client.chat.completions.create(model=deployment,messages=[{"role":"system","content":"You are a text-analysis tool. For the text, give sentiment (positive/negative/neutral), the entities, the key phrases, and a one-sentence summary."},{"role":"user","content":review}]) - Then Use the answer — pull the text out of the response and print it:
print(resp.choices[0].message.content) - Run it. Your program just did all four of yesterday’s jobs on real text. Change
reviewto a different sentence and run again — the same few lines now analyse anything you feed them. That is a lightweight application: a small, reusable piece of code that does one useful job well.
Notice the shape you already know from Day 14: a system message setting the rules, a user message carrying the content, and
choices[0].message.content to read the reply. Same pattern, new job.
26:00 – 30:00
Fix it in your own words
- Point at each part of your code and name its step: which lines are Import, which are Connect, which are Send, which are Use. If you can label all four, you understand the whole program.
- In your AI-901 Notes, write: “A lightweight text-analysis app = Import the library → Connect with endpoint+key → Send a system+user prompt → Use the printed result.”
- Save your Colab notebook (File → Save). You built something today — keep it. You will reuse this exact skeleton for speech and vision later this week.
Keep your key out of any screenshot you send me. Blur or crop it first — protecting a secret in a screenshot is a habit real professionals are judged on.
Key Terms
Client
An object in your code that knows how to reach your deployed model — created once from your endpoint and key, then used to send every request.
Endpoint & Key
The web address of your model and the secret password that authorises requests to it. Both come from your Foundry project; the key must never be shared.
Lightweight Application
A small, focused program that does one useful job well — here, reading text and returning analysis — that you can reuse on any new input.
Import → Connect → Send → Use
The four-step pattern behind every program that talks to an AI service. Learn it once; it fits text, speech and vision alike.
Today’s Assignment
Build and run a working text-analysis program in Colab.
- Build the program following the four steps. Type every line yourself — do not copy-paste. Confirm in one line that it runs without errors.
- Run it on at least three different pieces of text (reuse your three from yesterday if you like). Paste the printed output for each into your email.
- Take a screenshot of your notebook with a result showing — with your key not visible (blur or crop the Connect cell).
- Write two or three sentences: which of the four jobs did the model do best, and did anything surprise you? If you hit an error you could not fix, describe it — reporting a blocker is part of the assignment, not a failure.
Submit assignment 17 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. In the four-step pattern, which step is
client = AzureOpenAI(azure_endpoint=endpoint, api_key=key, ...)?- A. Import
- B. Connect
- C. Send
- D. Use
B. Connect — Creating the client from your endpoint and key is the Connect step: you are wiring the line to your deployment. Import is the
from openai import ... line; Send is chat.completions.create; Use is reading and printing the reply.2. Where should your API key go when you build this app?
- A. In a screenshot you send to someone
- B. In a public GitHub repository so others can help
- C. In a variable in your own private notebook, never shared or shown publicly
- D. In the subject line of your assignment email
C. In a variable in your own private notebook — The key authorises real spending on the subscription, so it stays private: never in a screenshot, an email, or anything public. In real projects it is kept out of the code entirely; for learning, a private notebook variable is acceptable.
3. Your program reads
resp.choices[0].message.content. What is this doing?- A. Sending the prompt to the model
- B. Connecting to the endpoint
- C. Using the model’s reply — pulling the answer text out of the response
- D. Importing the library
C. Using the reply — That line reaches into the response object and pulls out the model’s answer text so you can print or use it — the Use step. The Send step was the
create(...) call that produced resp in the first place.Exam Objectives Covered
Foundry SDK (Python)Text-analysis applicationSystem & user promptsEndpoint & key handlingImport→Connect→Send→Use
Domain 2 · SpeechDAY 18
Speech: Recognition, Synthesis, and Responding to Spoken Prompts
Why This Matters
Text is powerful, but most people would rather talk. Speech AI is what lets a phone take a spoken question and answer out loud — and in a country with dozens of languages and many people who prefer voice to typing, that is enormous. Think of a farmer asking about crop prices by voice, or a clinic line that answers in the caller’s language. Today you learn the two halves of speech AI and how they join up with the language model you already know. This is the kind of capability that turns a good app into one people actually use.
30-Minute Module
0:00 – 7:00
The two halves: recognition and synthesis
- Speech recognition turns spoken audio into written text. It is also called speech-to-text (STT). You talk; it types. This is what powers voice notes that appear as captions and voice-typing on a phone.
- Speech synthesis does the reverse: it turns written text into spoken audio. It is also called text-to-speech (TTS). You give it words; it reads them aloud in a natural-sounding voice.
- Learn both name-pairs cold for the exam: recognition = speech-to-text = STT; synthesis = text-to-speech = TTS. Examiners swap the terms freely, so you must recognise all of them.
- In Azure these live in the Azure AI Speech service — a prebuilt capability, like Day 16’s Language service. You send audio and get text (STT), or send text and get audio (TTS). You do not train anything; the models are ready-made.
Recognition listens (speech → text); synthesis speaks (text → speech). If you remember “recognise what I said, synthesise a reply”, you will never mix them up.
7:00 – 17:00
Joining the pieces: a spoken assistant
- Here is the beautiful part — the two speech halves sandwich the language model you already deployed. A voice assistant is just three steps in a row: (1) speech-to-text turns the caller’s spoken question into text; (2) your language model reads that text and writes an answer (exactly Day 17’s Send step); (3) text-to-speech reads the answer back aloud.
- Draw it: voice in → STT → text → language model → text → TTS → voice out. Every voice assistant you have ever used is some version of that loop.
- Notice this reuses everything: the system/user prompts from Day 12, the deployment from Day 13, the Import→Connect→Send→Use pattern from Day 17. Speech does not replace what you learned — it wraps around it.
- Two extra ideas the exam likes: some speech services also do translation (recognise in one language, produce text or speech in another), and real-time recognition can happen as you speak (streaming) rather than only after you finish.
A voice assistant is not one magic model — it is STT + a language model + TTS, chained. Seeing it as three familiar steps takes the mystery out of it and is exactly how the exam frames it.
17:00 – 25:00
See it work — without any setup
- You do not need to wire up microphones today to grasp this — you already use both halves daily. Find the voice-typing button on your phone keyboard (usually a small microphone) and dictate a sentence. Watch your speech become text — that is speech recognition in your hand right now.
- Now have your phone read something aloud — most phones have “Speak selection” or a read-aloud option in accessibility settings, or a maps app giving directions. That voice is speech synthesis.
- In your Colab notebook, sketch the assistant loop in comments only (no need to run it) so the shape lives in your fingers:
# 1. speech_to_text(audio) -> user_text# 2. send user_text to language model -> answer_text# 3. text_to_speech(answer_text) -> spoken audio - The point today is understanding, not a working microphone app. If you can explain the three-step loop and point to STT and TTS on your own phone, you have the exam-level grasp AI-901 asks for.
You are already surrounded by speech AI — voice-typing and read-aloud are STT and TTS. Naming the technology you already use is half the exam.
25:00 – 30:00
Fix it in your own words
- Say the two pairs aloud: recognition = speech-to-text = STT; synthesis = text-to-speech = TTS. Then say the three-step assistant loop from memory.
- In your AI-901 Notes, write one voice-assistant idea that would genuinely help someone where you live — a spoken crop-price line, a clinic booking line, a market-price checker — and note which step (STT, language model, TTS) does what.
- Tomorrow you move from ears to eyes: computer vision. The same three-in-a-row thinking will help you there too.
Speech AI is where accessibility and real-world use meet. An app that listens and speaks reaches people that a typing-only app never will — keep that value in mind.
Key Terms
Speech Recognition (STT)
Turning spoken audio into written text. Also called speech-to-text. Powers voice-typing and live captions.
Speech Synthesis (TTS)
Turning written text into spoken audio in a natural voice. Also called text-to-speech. Powers read-aloud and voice assistants’ replies.
Azure AI Speech
The prebuilt Azure service that provides speech-to-text, text-to-speech and speech translation — ready-made, no training required.
Voice Assistant Loop
The three-step chain: speech-to-text → language model → text-to-speech. Voice in, understood, answered, spoken back out.
Today’s Assignment
Explain speech AI in your own words and design a voice assistant for your community.
- In your own words (three or four sentences), explain the difference between speech recognition and speech synthesis, giving both alternative names for each (STT / TTS).
- Try both on your phone: dictate a sentence with voice-typing (recognition) and have your phone read a sentence aloud (synthesis). Confirm in one line that you did both and what you noticed.
- Design a voice assistant that would help people where you live. Describe the spoken question a user asks, then walk through all three steps of the loop — STT, language model, TTS — saying what each one does with the request. Paste the three-step comment sketch from your notebook.
Submit assignment 18 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. A phone app turns a user’s spoken question into text so it can be processed. What is this called?
- A. Text-to-speech (synthesis)
- B. Speech-to-text (recognition)
- C. Sentiment analysis
- D. Image generation
B. Speech-to-text (recognition) — Turning spoken audio into written text is speech recognition, also called speech-to-text or STT. Text-to-speech is the reverse — reading text aloud.
2. In the three-step voice assistant loop, what is the job of the middle step?
- A. To turn the user’s voice into text
- B. To read the answer aloud
- C. To read the recognised text and write an answer (the language model)
- D. To translate the audio into an image
C. The language model writes the answer — The loop is STT → language model → TTS. The first step recognises speech into text, the middle model reasons and replies in text, and the last step speaks the reply. The middle step is your Day-13 deployment doing its Day-17 job.
3. Which Azure service provides both speech-to-text and text-to-speech?
- A. Azure AI Speech
- B. Azure AI Language
- C. The model catalog
- D. The playground
A. Azure AI Speech — Speech recognition, synthesis and translation are all part of the prebuilt Azure AI Speech service. Azure AI Language handles text jobs like sentiment and entities; the catalog and playground are Foundry features, not speech services.
Exam Objectives Covered
Speech recognition (STT)Speech synthesis (TTS)Azure AI SpeechSpeech translationResponding to spoken prompts
Domain 2 · VisionDAY 19
Computer Vision: Interpreting Images With a Multimodal Model
Why This Matters
Today AI gets eyes. A model that can look at a photo and tell you what is in it opens doors everywhere — reading a receipt, checking whether a delivered product matches the order, describing a scene for someone who cannot see it, sorting photos automatically. You already know how to send text to a model; today you learn it can take an image too. The word for that is multimodal, and it is one of the most valuable ideas on this exam. Understand it well — “an AI that can see” is a phrase that gets you hired.
30-Minute Module
0:00 – 6:00
What computer vision and “multimodal” mean
- Computer vision is the branch of AI that interprets images and video — getting meaning out of pictures the way NLP gets meaning out of text.
- A modality is a type of data: text is one modality, images another, audio another. A multimodal model can handle more than one type — for example, take an image and answer in text. “Multi” = many; “modal” = types.
- This is a big shift from Week 2, where models only read and wrote text. A multimodal model you deploy in Foundry can look at a photo you send and describe it, answer questions about it, or read the text inside it.
- You met this word already: on Day 9 and Day 11 the model catalog labelled some models “multimodal”. Today that label finally means something concrete — those are the models with eyes.
Modality = a type of data (text, image, audio). Multimodal = one model that handles several. An image-and-text model is the classic example the exam uses.
6:00 – 16:00
What vision models can actually do
- Image description / captioning — the model writes a sentence saying what is in a picture: “a man selling oranges at a roadside stall”.
- Object detection — it finds specific items and where they are: “three phones and a charger on a table”. Classification is the simpler cousin: putting a whole image into a category (“this is a receipt”, “this is a passport”).
- Optical character recognition (OCR) — reading printed or handwritten text out of an image. Point it at a photo of a receipt and it returns the words and numbers as text you can use. This one is quietly one of the most useful in real business.
- A modern multimodal language model rolls many of these into one: you send an image plus a question (“how much is the total on this receipt?”) and it answers in plain language. That is Day 12’s prompting, now with a picture attached.
Four vision jobs to know: describe/caption, detect/classify objects, and read text (OCR). A multimodal model can do several at once when you ask it in a prompt.
16:00 – 26:00
Send an image to a model — the shape of it
- The four-step pattern still holds: Import → Connect → Send → Use. The only thing that changes is what you Send — now the user message carries an image as well as text.
- You need a multimodal deployment for this. If your Day-13 deployment is text-only, that is fine for understanding today; note that in a real build you would deploy a multimodal model from the catalog (inside your own resource group, as always).
- The user message becomes a small list of parts — some text, one image. In Colab, sketch the shape (you can run it if you have a multimodal deployment and an image URL):
messages=[{"role":"user","content":[{"type":"text","text":"What is in this image? Read any text you see."},{"type":"image_url","image_url":{"url":"https://.../receipt.jpg"}}]}] - Then Send and Use exactly as before:
client.chat.completions.create(model=deployment, messages=messages)and readresp.choices[0].message.content. The reply is text describing your image. The pattern you learned Tuesday just grew an extra sense.
Sending an image is not a new program — it is the same Send step with an image part added to the user message. If the four-step pattern is solid, vision is a small extension, not a fresh start.
26:00 – 30:00
Fix it in your own words
- Define aloud: modality, multimodal, computer vision, OCR. If you can say all four cleanly, you have today’s core.
- In your AI-901 Notes, write one real task near you that OCR-plus-a-question would solve — reading receipts for a shop, checking IDs, digitising handwritten records — and note that it is a multimodal model doing it.
- Tomorrow is the flip side: instead of understanding an image, you generate one. Same week, opposite direction.
If “multimodal” still feels abstract, anchor it: your phone camera that can already copy text out of a photo is OCR in a multimodal tool. You have seen this — now you can name it.
Key Terms
Computer Vision
The branch of AI that interprets images and video — getting meaning out of pictures, as NLP does out of text.
Multimodal / Modality
A modality is a type of data (text, image, audio). A multimodal model handles more than one — e.g. takes an image and answers in text.
OCR
Optical character recognition — reading printed or handwritten text out of an image and returning it as usable text.
Object Detection & Classification
Detection finds specific items and where they are in an image; classification sorts a whole image into a category.
Today’s Assignment
Show you can make a model interpret an image — in words, and in the shape of the code.
- In your own words (three or four sentences), explain what “multimodal” means and why a multimodal model is more useful than a text-only one. Give one example of each of the four vision jobs (caption, detect/classify, OCR).
- Write the message structure that sends an image plus a question to a model (the
contentlist with a text part and an image part). If you have a multimodal deployment, run it on a real image and paste the reply; if not, paste your sketched code and explain what each part does. - Pick one real image task from your area (reading receipts, checking a delivered product matches the order photo, describing scenes for a blind user). Describe the image you would send, the question you would ask, and the answer you would expect. Confirm in one line which resource group this would run in.
Submit assignment 19 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. A model can take an image as input and answer questions about it in text. What word describes this model?
- A. Text-only
- B. Multimodal
- C. Extractive
- D. Zero-shot
B. Multimodal — Handling more than one type of data (here image in, text out) is what “multimodal” means. A text-only model cannot take images; extractive and zero-shot describe summarisation and prompting, not model inputs.
2. You photograph a receipt and want the model to return the printed numbers as text. Which capability is that?
- A. Speech synthesis
- B. Sentiment analysis
- C. Optical character recognition (OCR)
- D. Image generation
C. OCR — Reading text out of an image is optical character recognition. Speech synthesis reads text aloud, sentiment judges tone, and image generation creates pictures — none of those extract text from a photo.
3. To send an image to a chat model, what changes compared with a text-only request?
- A. You must abandon the four-step pattern entirely
- B. The user message carries an image part alongside the text part
- C. You no longer need an endpoint or key
- D. The model stops needing a deployment
B. The user message carries an image part alongside the text — It is the same Import→Connect→Send→Use pattern; only the Send content grows an image part. You still need a deployment (a multimodal one), an endpoint and a key.
Exam Objectives Covered
Computer visionMultimodal modelsImage descriptionObject detection / classificationOCR
Domain 2 · Generative VisionDAY 20
Image Generation: Creating Visual Outputs With Generative Models
Why This Matters
You close Week 4 with the most eye-catching skill of all: making pictures from words. Yesterday a model read images; today one creates them from a text description you write. For a small business this is real money saved — product mock-ups, social-media posts, logos, ad images, all from a good prompt instead of a design budget. And it brings your prompt-engineering skill full circle: the quality of the picture depends entirely on how well you describe it. You are almost through the syllabus, both of you — one week to go. Finish this one strong.
30-Minute Module
0:00 – 7:00
What image generation is
- Image generation is a generative AI capability (Day 6) where a model creates a brand-new image from a text description. The text you write is called the prompt, exactly as with a language model — only the output is a picture, not words.
- This is the mirror image of yesterday. Day 19: image in, text out (understanding). Day 20: text in, image out (creation). Same idea — crossing between modalities — running the other way.
- The models are generative, so like a language model at higher temperature they produce something new each time — ask twice and you get two different pictures from the same prompt. There is no single “correct” image, just what your description steers it toward.
- In Azure you reach image-generation models (such as the DALL-E family) through the same Foundry catalog and deployment flow you already know — deployed, as always, inside your own resource group against your subscription.
Text in, image out. It is generative AI — new every time — and it uses the very same prompt skill from Day 12, just aimed at a picture instead of a paragraph.
7:00 – 17:00
Prompting for images: describe like a director
- A weak image prompt gets a weak image. “A shop” gives you something generic; a strong prompt describes subject, setting, style, and detail: “a bright, modern phone-repair shop in Lagos, wooden counter, shelves of devices, warm daylight, photographic style”.
- Four things worth naming in an image prompt: (1) the subject (what is in it), (2) the setting (where, time of day, mood), (3) the style (photo, cartoon, watercolour, logo), (4) detail (colours, camera angle, what to include or avoid). This is the visual version of Day 12’s “role, task, format, limits”.
- The Send step in code mirrors what you already know — a different method, same shape:
img = client.images.generate(model="my-image-model", prompt="a modern phone-repair shop in Lagos, warm daylight, photographic", n=1)print(img.data[0].url) - The result comes back as a link (a URL) to the generated image, which you open in a browser. Import → Connect → Send → Use, one more time — you Use the result by opening the picture.
Describe like a film director briefing a set: subject, setting, style, detail. The richer and clearer your words, the closer the image lands to what is in your head.
17:00 – 25:00
Using it well — and using it responsibly
- Real uses for a small business: product mock-ups before anything is made, social-media and advert images, simple logos and banners, illustrations for a website. One person with good prompts replaces a chunk of a design budget — that is a service you could offer.
- Now bring back Week 2’s Responsible AI. Generated images raise real questions: they can produce misleading or fake pictures (a reason transparency matters — be honest that an image is AI-made), and they must respect people and copyright. Do not generate images of real people in false situations.
- Cost awareness again (Day 9): each generated image spends a little real credit on your subscription. Generate with intent — write a good prompt and make it count — rather than clicking dozens of times to see what happens.
- And the key rule holds through every capability this week: your endpoint and key authorise this spending, so they stay private. A leaked key could run up a bill on my account overnight.
The technology is exciting, but the professional habit is what gets you hired: generate with purpose, be transparent that an image is AI-made, and respect real people and copyright.
25:00 – 30:00
Fix it in your own words — and look back at your week
- Say aloud the pair that defines this week’s two vision days: Day 19 was image → text (understanding); Day 20 is text → image (creation). Both are multimodal, in opposite directions.
- In your AI-901 Notes, write the four parts of a strong image prompt (subject, setting, style, detail) and one responsible-AI caution about generated images.
- Step back: this week you did text analysis, built an app, learned speech, and handled vision both ways. That is the whole “capabilities” span of Domain 2. Next week is information extraction and exam prep — the home straight. Be proud of how far you have come from Day 1.
Reaching the end of Week 4 from zero technical background is a real achievement — not everyone who starts gets here. One week left. Keep the same steady pace and you will meet that exam ready.
Key Terms
Image Generation
A generative AI capability that creates a brand-new image from a text prompt. Text in, image out — new every time you run it.
Image Prompt
The text description you give an image model. Strong ones name subject, setting, style and detail — the visual version of good prompt engineering.
Generative Model
A model that produces new content rather than classifying existing content. Image generators, like language models, create something fresh each time.
Responsible Generation
Using image generation honestly: disclosing AI-made images, avoiding misleading or fake depictions of real people, and respecting copyright.
Today’s Assignment
Write image prompts like a director — and reflect on your whole week.
- Write three strong image prompts for a real small business near you (a product mock-up, a social-media post, a logo or banner). For each, point out the four parts: subject, setting, style, detail.
- Take one weak prompt — “a shop” — and rewrite it into a rich one, then explain in a sentence what you added and why it will improve the picture. Paste the
images.generatecode shape you would use, and confirm which resource group it runs in. - Write two or three sentences on responsible use: name one risk of generated images and how you would handle it honestly.
- Finish with one line looking back: across this week — text, speech, vision — which capability excites you most for building something real, and why?
Submit assignment 20 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. Image generation is best described as which kind of task?
- A. Image in, text out (understanding an image)
- B. Text in, image out (creating a new image)
- C. Audio in, text out
- D. Sorting an image into a category
B. Text in, image out — Image generation creates a new picture from a text prompt. Option A describes yesterday’s computer vision, C is speech recognition, and D is image classification — all different tasks.
2. You run the same image prompt twice and get two different pictures. Why?
- A. The service is broken
- B. Image models are generative and produce new output each time
- C. Your key changed between the two runs
- D. The prompt was too specific
B. They are generative — Like a language model producing fresh text, an image generator creates something new on each run, so the same prompt yields different pictures. That is expected behaviour, not a fault, and nothing to do with your key.
3. Which is a responsible-AI concern specific to generated images?
- A. They use tokens
- B. They can create misleading or fake pictures of real people, so AI-made images should be disclosed
- C. They must always be text-only
- D. They cannot be deployed in a resource group
B. They can be misleading, so disclose them — Generated images can depict people or events falsely, which is why transparency (being honest an image is AI-made) and respect for real people and copyright matter. This is Week 2’s Responsible AI applied to a new capability.
Exam Objectives Covered
Image generationGenerative modelsImage promptingResponsible AI (transparency)Capability vs cost