Part 4
What could you build
in your own role?
Needs differ by setting. Pick the closest example, follow it as written, then adapt it to your own work once it feels familiar.
Finding your example
Find the row closest to your situation and jump to that section. The other examples are worth reading too if you have time.
| If you are | Your usual problem | Go to |
|---|---|---|
| Faculty · researcher | Organizing papers and materials eats the day | 4.1 Organizing research material |
| Computing student | You want to build something and keep it as a portfolio piece | 4.2 A document Q&A assistant |
| Student in another field | Code is daunting but you want to use this for study | 4.3 A study material organizer |
| Nonprofit staff | Paperwork and reporting take most of your time | 4.4 Automating document checks |
| University hospital | You want to use it safely within the rules | 4.5 Forms and safety standards |
The problem → what you build → in what order → what you end up with. For the building itself, choose whichever suits you: Part 2 (no code) or Part 3 (code).
4.1Faculty and researchers
"Reading and organizing papers takes a whole day." Gathering and comparing material is where the time goes.
What you build
A tool that takes several paper PDFs and produces title, authors, method, key findings, and relevance to your topic in a consistent format. Collected into a table, they become easy to compare.
Convert the PDFs to text
Use Document Parse. For papers with many tables and figures, Enhanced mode is worth using.
Extract the same fields from each
Title, authors, year, method, sample size, key findings, limitations. Fixing the format means every paper comes out in the same shape.
Have it judge relevance to your topic
Ask something like "my research topic is X — rate how relevant this paper is as high, medium, or low and give one sentence of reasoning".
Collect into a table
Move the results into a spreadsheet to sort and compare. Read the highly relevant ones properly first.
You can use it for teaching too
| Use | How |
|---|---|
| Draft teaching materials | Feed in lecture notes and pull out learning objectives and key questions |
| Create practice problems | Generate variations of a problem at different difficulty levels |
| Review criteria for student work | Define assessment dimensions and require reasoning for each |
Treat the AI's judgment as reference material and keep the final assessment with a person. If you did use AI in assessment, it is good practice to tell students how.
Read Part 6 first if the material contains student personal data.
Extract the same fields across many documents and collect them into a table — reading turns into comparing.
4.2Computing students
"I want to build something and have it to show." You need experience understanding the structure and implementing it.
What you build
A document Q&A assistant: upload a PDF, ask questions about its contents. It touches document processing, structured output, and tool calling all at once, which makes it a good learning project.
Convert the document to text
Use Document Parse to get structured text from the PDF. Make use of the page numbers in elements[].
Answer questions from it
Send the document text along with the question. Requiring the page number it relied on makes the answers much more trustworthy.
Fix the output format
Define fields for the answer, the source page, and a confidence level, and displaying it becomes straightforward.
Let it say when it does not know
Add "if the answer is not in the document, say 'not found in this document' instead of guessing". That one line changes the quality substantially.
Where to take it further
| Extension | What you learn |
|---|---|
| Handle multiple documents | The problem of choosing which document holds the answer |
| Add tools | How to route questions requiring calculation or lookup to a function |
| Build an evaluation set | Prepare question-answer pairs and compare accuracy after each change |
How you solved the problem reads better than a working screen. Record what was wrong in your first version, what you changed, and how much it improved.
Keep your API key out of your code and in an environment variable. Keys end up in repositories more often than you would think.
One document Q&A assistant covers document processing, structured output, and tool calling. Record how you improved it.
4.3Students in other fields
"Code is daunting, but I want to use this for studying." You want something you can use right away.
What you build
A tool that takes lecture slides or a textbook chapter and produces key concepts, a summary, and practice questions. Two or three blocks in Studio (Part 2) are enough.
Upload the material
PDFs, photos of handwritten notes, and HWP files all work.
Decide how it should be organized
Write what you want, such as "pull out five key concepts and explain each in two sentences".
Have it write practice questions
Ask for "five multiple-choice questions with answers and explanations based on this material".
Other ways to use it
| Situation | Try asking |
|---|---|
| A concept you find difficult | "Explain this as you would to a high school student, with one analogy." |
| Skimming a long reading | "Give me one sentence per chapter as a list." |
| Exploring careers | "Describe what this role does and what skills it needs, and what I could start preparing now." |
Check your institution's or instructor's AI policy first. What is permitted and how you must disclose it varies from course to course.
Rather than submitting AI output directly, check it and rewrite it in your own words. It helps you learn, and AI can state things that are simply wrong.
Organizing material and generating questions work without any code. Just do not use the output without checking it.
4.4Nonprofit staff
"Document checks and reports take all my time." The smaller the team, the heavier repetitive paperwork feels.
What you build
A tool that takes applications or expense receipts and extracts the fields you need while flagging missing documents. It leaves you only the items that need a human.
Convert the documents to text
Korean HWP files can go in directly. Use OCR Force for scans.
Extract the fields you check every time
Organization, project name, amount requested, period, list of attachments.
Sort by your criteria
Something like "if all required documents are present, ACCEPT; if any is missing, NEEDS_REVISION; if the amount or organization cannot be found, NEEDS_CHECK".
Leave only what needs a person
Staff only look at NEEDS_REVISION and NEEDS_CHECK. There is no need to open every file.
The same approach fits these too
| Work | What to extract | What to judge |
|---|---|---|
| Expense receipts | Merchant · date · total · items | Limits, missing evidence |
| Donor reporting | Programme · spending · people reached | Difference from the plan |
| Meeting notes | Decisions · owners · deadlines | Flag anything past its deadline |
| Funding calls | Eligibility · deadline · required documents | Whether your organization qualifies |
Donor lists and beneficiary records often contain personal data. Read Part 6 · Using It Safely first, and mask names and contact details where needed.
Document Parse allows 1 request per second. Sending dozens at once will stall partway, so send them one at a time. See the loop example in Part 3.
For high-volume work, code is the better fit. Studio only covers Parse, while code covers Solar as well. Design it in Studio first, then move it over.
Instead of checking every document, filter down to the ones that need attention. The workload drops sharply.
4.5University hospitals
"We want to use it safely within the rules." Given how sensitive the material is, deciding what not to submit comes first.
For any material containing patient information, confirm your institution's data protection policy and review process first. Your institution's rules take precedence over this document.
The examples below assume material that contains no patient-identifying information.
What you build
Tools for organizing documents that contain no patient data — research protocols, conference material, administrative forms.
| Use | Material | What it does |
|---|---|---|
| Literature review | Published papers | Puts design, sample, and findings into a consistent format |
| Administrative forms | Application and internal forms | Checks for missing required fields |
| Teaching material | Lecture slides, guidelines | Summarizes key points and drafts questions |
| Policy comparison | Internal guidelines, notices | Finds what changed and links related clauses |
Four things that keep it safe
Decide what not to submit, first
Make it a rule not to include anything that identifies an individual — names, dates of birth, record numbers, contact details, diagnoses.
Mask before submitting
Where it is genuinely necessary, remove or replace identifying details first. Keep masked copies separate from originals.
A person reviews the output
Anything feeding into clinical work must be checked by the responsible person. AI output is reference material.
Keep a record
Noting what material was used for what purpose helps when it needs to be reviewed later.
What you enter in the Playground may be used to improve the model. Do not enter material containing patient information.
See Part 6 · Using It Safely for details.
Decide what not to submit before deciding what to build. Your institution's rules always come first.
4.6[Activity] Write your own
Pick the closest example and rewrite it for your situation. This single page is enough to start building in Part 2 or Part 3.
# My use case
## Example I am following
- (one of 4.1 – 4.5):
## My situation
- Organization:
- What takes the most time right now:
- How many per month:
## What I will build
- Put in:
- Get out:
## The material
- Format:
- Does it contain personal data (yes / no):
- If yes, how will you handle it:
## Three fields to extract
1.
2.
3.
## Criteria for judgment
- What conclusion is needed:
- The criteria, as numbers and conditions:
## Where a person checks
-
## How I will build it
- [ ] Part 2 (Studio, no code)
- [ ] Part 3 (code)Checkpoint
- You chose an example to follow.
- You wrote one sentence each for what goes in and what comes out.
- You checked for personal data and decided how to handle it.
- You listed three fields to extract.
- Your criteria are numbers and conditions.
- You identified where a person checks.
Begin with one document type and one judgment. Confirm it works, then widen the scope — that gets you there much faster.
Part 4 key takeaways
- The approach is the same everywhere: convert → extract → judge.
- Faculty and researchers gain most from collecting the same fields for comparison.
- Nonprofit staff can cut the load by filtering down to what needs attention.
- Students should check the output and rewrite it in their own words.
- University hospitals should decide what not to submit first, with institutional rules taking precedence.
Where to go next
If something goes wrong, Part 5 is organized by symptom. And please read Part 6 before working with real material.