LIN CHIACHENG

02 / 07GRADUATION PROJECT · TEAM OF 4

NOTERA AI

Class, video and image material, arranged into readable notes with AI.

Role

AI Document Intelligence · AI / Product Builder · graduation project team lead

Period
2025 - 2026
Status
Confirmed

01

IDEA

It began with one direction only.

What if the information in class, videos and images could be arranged into readable notes by AI?

Starting the work showed the hard part was not asking AI to summarize, but deciding what to keep, how images are understood, how code survives, and how a team finishes within the time.

IDEA FLOW

VIDEO / IMAGE / SLIDE → AI → NOTE
  1. 01

    Video / Image / Slide

    Video / image / slide

  2. 02

    AI

    AI reading

  3. 03

    Note

    Readable note

THREE PARTS

  1. 01VIDEO / IMAGE / SLIDE

    Video / image / slide

    Bring the class material in as it is.

  2. 02AI

    AI reading

    Read text together with the visuals and organize the content.

  3. 03NOTE

    Readable note

    Arrange it into structured notes.

CHAPTER NOTEOne direction was easy; deciding what goes inside it was the hard part.NOTERA / 01 · IDEA

02

SCOPE

Many wants, so choose first.

We split the work into WANT / MUST / LATER and first fixed a scope that could finish in the time we had.

We could not do everything, so we separated what stays from what waits.

SCOPE SHEET

WANT / MUST / LATER
WANTWANTED
  1. 01Class, video and image material arranged with AI
  2. 02Kept as readable notes
MUSTREQUIRED
  1. 01Image and video content can be processed
  2. 02Structured notes can be generated
  3. 03Multilingual and code content is not filtered incorrectly
  4. 04It can actually be demoed
LATERLATER
  1. 01More complete knowledge management
  2. 02A larger model / more complex analysis
  3. 03Further personalization

SCOPE RULE

WANT / MUST / LATER

  1. 01WANT

    What we wanted

    The wide starting direction; the cut starts here.

  2. 02MUST

    Conditions for finishing in time

    Image and video processing, structured notes, multilingual code content, and a demo.

  3. 03LATER

    What waits

    Three items placed outside the scope; return to them later.

CHAPTER NOTEMore features is not better; making the main flow actually run came first.NOTERA / 02 · SCOPE

03

TEAM

As team lead in a team of four, I helped set the direction and took part in the development.

As team lead, I helped shape the topic and overall architecture, then implemented my assigned parts; teammates supported research, the report and the presentation.

Not everyone needed to do the same work from the start; we first put into words what to build and how far to go.

RESPONSIBILITY MAP

TEAM OF FOUR

My responsibilities as team lead

  1. 01IDEATopic proposal
  2. 02ARCHITECTUREOverall architecture
  3. 03INTEGRATIONIntegration calls

Teammates’ responsibilities

  1. 01DOCUMENTMaterial and reports
  2. 02TESTDelivery checks
  3. 03PRESENTATIONPPT and presentation

WHAT EACH SIDE DID

DEVELOPMENT / MATERIAL

  1. 01My responsibilities as team lead

    IDEA / Topic proposal

    Frame the question and set the main concept.

  2. 02My responsibilities as team lead

    ARCHITECTURE / Overall architecture

    Decide the design and development direction.

  3. 03My responsibilities as team lead

    INTEGRATION / Integration calls

    Align the joins and make the final calls.

  4. 04Teammates’ responsibilities

    DOCUMENT / Material and reports

    Organize the material into the report.

  5. 05Teammates’ responsibilities

    TEST / Delivery checks

    Support the checks before delivery.

  6. 06Teammates’ responsibilities

    PRESENTATION / PPT and presentation

    Prepare the PPT and presentation material.

CHAPTER NOTEMy job was not to finish everything alone, but to fix the topic and architecture first and point four people's work at the same outcome.NOTERA / 03 · TEAM

04

PLAN

Cut the work to a size that finishes in time.

We did not each start writing our own part first; we confirmed how the material comes in, how it is processed, and how it is finally seen.

PROJECT TIMELINE

RESEARCH → UI → CORE → DOCS → PRESENTATION

  1. 01RESEARCH

    Check how material enters.

    OWNER
    My responsibilities as team lead
    OUTPUT
    How material comes in

  2. 02UI

    Implement how it is shown.

    OWNER
    My responsibilities as team lead
    OUTPUT
    How it is seen

  3. 03CORE

    Connect the whole.

    OWNER
    My responsibilities as team lead
    OUTPUT
    The joined main flow

  4. 04DOCS

    Organize material and the report.

    OWNER
    Teammates’ responsibilities
    OUTPUT
    The technical report

  5. 05PRESENTATION

    Prepare the PPT and presentation material.

    OWNER
    Teammates’ responsibilities
    OUTPUT
    The PPT and presentation material

THREE QUESTIONS

IN / PROCESS / VIEW

  1. 01IN

    How material comes in

    Fix the input shape first.

  2. 02PROCESS

    How it is processed

    The reading and arranging path.

  3. 03VIEW

    How it is finally seen

    Include the notes view in the plan.

CHAPTER NOTEFixing the order kept the work moving inside a fixed timeline.NOTERA / 04 · PLAN

05

BUILD

The flow from input to notes.

What we built is one flow: INPUT → UNDERSTAND → ORGANIZE → OUTPUT.

Rather than listing technology names, we made where things connect visible.

BUILD FLOW

INPUT → UNDERSTAND → ORGANIZE → OUTPUT
  1. 01

    INPUT

    Video / image input

  2. 02

    UNDERSTAND

    Text and visual recognition

  3. 03

    ORGANIZE

    Summarize and de-duplicate

  4. 04

    OUTPUT

    Structured notes

Vue 3 / FastAPI / Local AI / Docker

FOUR STEPS

  1. 01INPUT

    Video / image input

    Load video and images as they are.

  2. 02UNDERSTAND

    Text and visual recognition

    OCR and a VLM (vision-language model) read the text and visuals together.

  3. 03ORGANIZE

    Summarize and de-duplicate

    Choose what is kept.

  4. 04OUTPUT

    Structured notes

    Arrange it into a readable form.

CHAPTER NOTEOnce the joins were visible, it was clear where the flow was stuck.NOTERA / 05 · BUILD

06

DECISIONS

From OCR-first to VLM-First; audio becomes slides only.

What mattered was not what we ended with, but how we changed course midway.

We replaced the center of the reading and narrowed the material; both moves protected the main flow within the limited time.

TWO TURNS

OCR-first → VLM-First

Audio + slides → slides only

DECISION 01

OCR-first → VLM-First

BEFORE

  1. OCR
  2. language/text decision
  3. notes

AFTER

  1. Screenshot + OCR
  2. VLM analysis
  3. notes

Code, English and mixed-language content was easily dropped by mistake.

WHY IT MOVED

OCR-first → VLM-First

  1. 01

    What was wrong

    Code, English and mixed-language content was easily lost.

  2. 02

    What changed

    Analyze screenshot plus OCR with a VLM, reading text and visuals together.

DECISION 02

Audio + slides → slides only

BEFORE

  1. Audio + slides
  2. decision

AFTER

  1. Slides
  2. output

The recorded audio was poor and did not work.

WHY IT MOVED

Audio + slides → slides only

  1. 01

    What was wrong

    We tried to use audio and slides together, but the audio quality was bad.

  2. 02

    What changed

    We changed to slides-only output.

CHAPTER NOTEChoosing to make the main path run first was the biggest decision in the project.NOTERA / 06 · DECISIONS

07

RESULT

Close with the real demo.

The actual screens show video input, processing status and the resulting notes.

The project was presented at graduation and reached a working demo. It runs qwen3-vl:4b locally through Ollama alongside PaddleOCR-VL. Short videos can be processed; longer videos are limited by the available local compute.

DEMO PLATES

INPUT / VIDEO LIST / NOTE
The video list and processing status screen
01 / INPUTThe video list shows pending and completed processing states.Record status · Confirmed
The loaded video list
02 / VIDEO LISTThe loaded video list, showing processing progress and storage locations.Record status · Confirmed
The generated lecture notes and key-term list
03 / NOTEThe processing result becomes lecture notes and a key-term list.Record status · Confirmed

WHAT REMAINS

Confirmed

  1. 01SUPPORTED

    Input through structured notes

    Joined through Vue 3, FastAPI and Docker.

    Confirmed

  2. 02SUPPORTED

    The four-person split and integration calls

    Cut the scope, chose the reading direction, and joined at the seams.

    Confirmed

  3. 03SUPPORTED

    The graduation-project presentation

    The presentation slides and technical report were completed and presented.

    Confirmed

Record status

NOTERA / 07

The close centers on the verified path from input to structured notes and the team's decisions.

CHAPTER NOTEMore than the screens, the flow from input to notes is what this project produced.NOTERA / 07 · RESULT

NOTERA AI — LIN CHIACHENG