Lab 2: AI Workload Explorer

Get hands-on with every Domain 1 workload in one sitting: run sentiment, key phrase and entity detection on customer reviews, convert speech to text and back, caption an image, read text from a scanned document, and finish the workload-matching sheet.

Time: 75–90 minutes · You need: the Foundry project from Lab 1 (with a multimodal chat model deployed, such as gpt-4o-mini or a GPT-5 family model), a short audio clip, an image and a scanned document (photos from your phone are fine).

Outcome. Hands-on familiarity with every AI workload tested in Domain 1.

Foundry gives you two ways to try each workload. The Foundry Tools playgrounds (Language, Speech and others, found under the tools / AI services area of your project) call the specialised services. The chat playground with a multimodal model does the same tasks through prompting. Try both where you can and compare the outputs.

Step 1: Text analysis on customer reviews

Sample reviews:

1. The new Contoso credit card has excellent cashback but the mobile app keeps logging me out.
2. Very disappointed. My UPI refund of Rs 2,450 from 12 March is still pending. Ref: TXN99812.
3. Ananya at the Koramangala branch was extremely helpful in opening my daughter's account.
4. Fees are okay. Nothing special.

A. With Azure Language (Foundry Tools playground). Open the Language playground in your project, choose each feature in turn, paste the reviews and run:

• Sentiment analysis: note the document label and the sentence-level labels. Review 1 should come out as mixed.
• Key phrase extraction: which phrases appear for review 2?
• Named entity recognition: find the Person, Location, DateTime and Quantity entities.
• PII detection: see the redacted version of review 2.

B. With the chat model. In the chat playground, try:

For each review below, return JSON with: sentiment (positive/negative/neutral/mixed),
key_phrases (list), entities (list of {text, category}). Reviews:
<paste the four reviews>

Compare: the Language service gives calibrated confidence scores and a fixed schema; the LLM is flexible but its output format and labels depend on your prompt.

Step 2: Speech to text and text to speech

1. Record a 10–15 second clip on your phone, for example: “Hi, I want to know why my debit card was charged an annual fee of four hundred rupees.” Save it as WAV or MP3.
2. In the Speech playground, open real-time speech to text, upload the file (or use your microphone) and read the transcript. Try the language set to English (India).
3. Open text to speech (voice gallery). Type a reply: “We have reversed the annual fee. It will reflect in two working days.” Try two voices, such as en-IN-NeerjaNeural and en-IN-PrabhatNeural, and change the speaking rate.

Step 3: Analyse and caption an image

In the chat playground (multimodal model), attach a photo, for example a shop front, a product or a street, and ask:

1. Write a one-sentence caption for this image.
2. List the main objects you can see.
3. Is there any visible text? Transcribe it exactly.

Then ask a follow-up visual question, such as “How many people are in the photo?” This is visual question answering.

Step 4: OCR on a scanned document

Attach a photo of a printed receipt, bill or form and ask:

Extract all printed text exactly as it appears, preserving line breaks.
Then list: merchant name, date, total amount.

Check what it got right and what it misread (smudges, small fonts, handwriting). In Session 10 you will do this at scale with Content Understanding, which returns typed fields with confidence scores.

Step 5: Workload-matching sheet

Pick the right Azure capability for each scenario.

Check yourself