Timetable Image AI Recognition and Prompt Improvement

Problem

For the Kakao Tech Campus final project, we built a student schedule management service. Manually entering timetables is tedious, so I wanted to auto-recognize them from uploaded images.

---

Approach: GPT Vision + Structured Output

I chose GPT-4.1’s Vision feature over traditional OCR. Why:

• Timetable formats vary wildly (Everytime app, school websites, handwritten, etc.)
• I needed structured data, not just raw text

I used GPT’s Structured Output to enforce the output format via JSON Schema.

---

Prompt Design

At first, I went simple: “Convert this timetable image to JSON.” Results were all over the place. Maybe 2–3 accurate out of 10 tries.

After refining the prompt, I hit 8–9 accurate results out of 10.

Final Prompt Rules:

1. dayOfWeek starts at 1 for Monday. (Mon=1, Tue=2, Wed=3, Thu=4, Fri=5)

2. startTime, endTime use "HH:MM:SS" format, 24-hour clock.

3. Same subject + professor + room should be grouped as one subject,
   with multiple time entries in the times list.

4. If credit isn't specified:
   - Regular lecture courses: 3
   - Lab/practice courses ('lab', 'practice', 'project', 'capstone', etc.): 2

5. If time isn't directly shown in the image, estimate based on grid spacing.

6. Infer class duration from the table. Don't assume 1 hour without evidence.

7. Start/end times may not be on the hour. Could be 5-minute intervals.

8. Enter all strings exactly as shown in the image.

---

Improvement Points

Problem	Solution
Always assumed 1-hour classes	Added “Don’t assume 1 hour without evidence”
Only guessed times on the hour	Added “Could be 5-minute intervals”
Same course split into multiple entries	Added “Group same courses” rule
Returned 0 credits	Added default credit rules by course type

---

Lessons Learned

• AI output varies a lot depending on the prompt
• Structured Output keeps parsing stable
• Anticipate edge cases and bake them into the prompt

---

From Kakao Tech Campus 3rd cohort final project (student schedule management service).

• kakao-tech-campus-3rd-step3/Team12_BE