On this page

Practice set · Block 1 · Week 2 · Building APIs and the backend

Building APIs

One chat app. One raw POST request. Proof against your golden pair.

Building APIs and the backendSelf-pacedBlock 1 · Ship the first system

The exercise in one sentence

Connect your Gradio app with a raw requests.post to an LLM, carrying your own prompt and your user's real input, and prove the reply is useful by checking it against your golden pair.

0/5

The five stepsSection 02

0/4

Prediction logThree requests, two failures

0/6

DeliverablesSection 07

Which track are you on?

01Why This Exercise Exists

A passenger enjoys the comfort of a car. A driver understands its controls and knows what to do when things go wrong.

Comfort without understanding comes at the cost of control.

In software, SDKs, libraries, no-code/vibe-coding tools are like comfortable passenger seats. They make the journey easier by hiding the machinery underneath.

But hidden machinery doesn't disappear.

When something breaks, what you understand determines what you can fix.

To drive software machines, understand its controls: APIs, endpoints, authentication, requests, and responses.

Use abstractions for comfort. Build understanding for control.

A 200 proves the call worked. Only your verifier proves the reply helps.

02The Five Steps

Track B · Aarav as the user

Input: his weekly workflow. Output: a plan to automate part of it. Starting prompt:

"Here is the workflow of an analyst working at an IT company in India. Diagnose and generate a plan to automate this work, or a part of the work."

Target: the plan drifts into jargon Aarav cannot use.

1. Start from the echoRun your Gradio chat from The first interface. Confirm it returns your input unchanged: no process yet.
2. Test the machine before the codeIn httpie.io/app, send a POST to https://api.groq.com/openai/v1/chat/completions with your key and a body naming openai/gpt-oss-20b. Write the status code you expect before you press send. Practise it on this page first.
3. Write your promptOne instruction that encodes what you learned doing your user's process by hand.Start from the prompt above. Send a prompt plus your user's real input in the browser first and read the reply as your user would.
4. Rip off the cushionIn Python, replace the echo with requests.post, using the same address, action, key and package. Return only the content field. No Groq or OpenAI SDK in the final file.
5. VerifyRun your golden input through the app. Compare the reply to your golden output, write a one-line verdict, change the prompt once based on what the comparison showed, and run it again.Find the jargon in the plan. Fix it with a prompt (no jargon, one step at a time, the minimum action needed now) and show the before and after. Use the verification note.

Steps done

0 / 5

Step 1 · Where you start

Step 4 · The shape of the change

Each blank is one of the four parts, with the values from the raw request. Write it yourself.

Your machine

export GROQ_API_KEY=gsk_your_key_here

$env:GROQ_API_KEY="gsk_your_key_here"

03The Raw Request, Live

Step 2's request, sent from this page to Groq. Pick your predicted status code to unlock Send; every send goes to the prediction log. Your screenshot still comes from httpie.io/app.

POST · Groq chat completionsFree tier: 30 requests a minute

Address

https://api.groq.com/openai/v1/chat/completions

Action

POST · a Create: each call creates a new reply

Key

Sent as Authorization: Bearer <key>. Held in this tab's memory only: never saved, gone when you reload. Get a key.

Package

model Your prompt · one instruction Your user's real input

The package on the wire

Break one part on purpose for this send:

What status code do you expect?

Pick a prediction first.

—

The full reply package

What each code means here

200	Groq understood you and sent a reply. It does not mean the reply helps anyone: that is section 05.
400	The package is wrong: broken JSON or a missing field.
401	The key is missing, mistyped, or sent without `Bearer`.
404	The model named in the package does not exist, or the address is wrong. Read the message: it says which.
429	Too many requests this minute on your account. Wait and send again.
5xx	Their machine, not yours. Try again later.

Status code says whose fault; the message says what. Try the fix the error points to before pasting anything into an LLM. More in Troubleshooting.

Got an error? Search it in Troubleshooting: the part it points to and the fix.

04Prediction Log

Three requests minimum, two of them deliberate failures. Sends from section 03 land here; add the rest by hand.

Three requestsA wrong model nameA missing or wrong keyA prediction on every row

Request	Predicted	Actual	Error message, for failures

05Verify Against Your Golden Pair

Golden output beside the reply, a one-line verdict, one prompt change, run again.Load Aarav's golden pair below. The target is jargon.

Golden inputWhat your user typed, word for word. Golden outputWhat a useful reply looks like, written by you by hand.

Actual output, before the changeThe reply your app returned. Actual output, after the changeThe reply once the prompt was changed.

One-line verdictDoes the actual output do what the golden output does, for this user? The one prompt changeWhat you changed, and which part of the comparison made you change it.

Saves on this device as you type.

06Rules of the Build

Predict before you run.

Every request in the prediction log gets its expected status code written down first. Include at least two deliberate failures: a wrong model name and a missing or wrong key. Record what each actually returned.

Read the error before you ask.

Status code says whose fault; the message says what. Try the fix the error points to before pasting anything into an LLM.

Minimum path.

No extra parameters, no chat history handling, no changed interface, no renamed functions. The diff from the echo app should be the process and nothing else.

Keys never in code or screenshots.

Set it in the terminal, never in the file. Blur the key in every image you share. If a key leaks, revoke it and create a new one: Keys.

Stay inside the free tier.

Groq allows 30 requests per minute. You will not need more.

LLM help: hints and debugging yes; generating the solution no.

Think for the first minute on your own. Pasting this brief in and copying the code out defeats the exercise; the code is short enough that writing it is the learning.

07Your Deliverables

App screenshotThe Gradio chat running locally, with your user's real input and the model's reply visible. Key not visible.
Code fileOne Python file (repo link or paste) using requests.post, with the address, action, key and package. No SDK import.
Raw request screenshothttpie.io/app showing POST, the full endpoint, and a 200 response, key blurred.
Prediction logAt least three requests (two deliberate failures), each with predicted status code, actual status code, and the error message for failures. Copy it from section 04.
Verification noteGolden output beside actual output, a one-line verdict, the one prompt change you made, and the reply after the change. Copy it from section 05.
The 201 questionTwo sentences: why did the chat completions call return 200 and not 201?

Deliverables ready

0 / 6

The 201 question

Why did the chat completions call return 200 and not 201?Two sentences, in your own words. A starting point: what a 201 meant on Aarav's service in The second interface, and whether Groq kept anything after it replied.

08Stretch: Any Machine, Four Lines

Pick one free API from the public-apis list (MediaStack is a fine start). Answer the four questions for it in writing before writing any code.
Call it with requests and print one field from the JSON response.
Optional swap test: point your chat app at another OpenAI-compatible endpoint by changing only the URL, key and model name. Note what else, if anything, had to change.

API Address Action Key Package Swap test

09How to Get Unblocked

Search the exact error in Troubleshooting first. Then compare your file, line by line, with the shape in section 02 and the package on the wire in section 03. Keep your deliverables together: they are your proof that the system works and helps.

Next: building your own API, and how much protection it actually needs.

Hold this

A 200 means the machine understood you. It does not mean you helped anyone. Four lines get you to any machine on the internet; your golden pair tells you whether the trip was worth it for a human.

Four lines to any machine.Companion pages: Overview · The second interface · Groq step · troubleshooting · back to the roadmap