Multimodal Engineering

Look at the attached image.
Step 1: Write all the text you can read, in order.
Step 2: Return JSON with these keys:
- item_type (for example receipt, sign, menu, note)
- title (the main heading, or null)
- dates (list)
- numbers (list of every number you can read, as written)
- unclear (list of parts you could not read well)
Do not guess. If a part is hard to read, put it in unclear and do not put it in the other keys.

Below is the text that came from a blurry photo of a shop receipt. Some characters are wrong.

---
RAHIM GR0CERY STOR3
12 Moy 2O25
Rice 5kg ....... 4SO
Oil 1L ......... 18O
Eggs (12) ...... ??5
Total .......... 8l5
---

Return JSON with these keys:
- item_type
- title
- dates (list)
- numbers (list of every number, as written)
- unclear (list of parts you could not read well)
Do not guess. If a part is hard to read, put it in unclear and do not put it in the other keys.

What do you think of this image?

This is a screenshot of an app screen.
1. Say in one sentence what the screen is for.
2. List up to 3 things that look wrong or confusing, such as cut-off text, low contrast, or unclear buttons. For each one, say where it is on the screen.
3. List what you cannot judge from this image.
Be specific. Do not praise the design.

Here is a description of a screenshot, because I cannot upload it.
The screen is a checkout page for a phone shop. At the top, a title says "Pay" in light grey on a white background. In the middle, a field named "Phone numbr" has a red border but no message. At the bottom, two buttons sit next to each other: a big green button with the text "OK" and a small grey button with the text "Cancel". The price "BDT 12,4" is cut off at the right edge.

What do you think of this screen?

Here is a description of a screenshot, because I cannot upload it.
The screen is a checkout page for a phone shop. At the top, a title says "Pay" in light grey on a white background. In the middle, a field named "Phone numbr" has a red border but no message. At the bottom, two buttons sit next to each other: a big green button with the text "OK" and a small grey button with the text "Cancel". The price "BDT 12,4" is cut off at the right edge.

1. Say in one sentence what the screen is for.
2. List up to 3 things that look wrong or confusing. For each one, say where it is on the screen.
3. List what you cannot judge from this description.
Be specific. Do not praise the design.

I build a voice agent for a bank help line. The design is speech-to-text (STT), then an LLM, then text-to-speech (TTS). After the caller stops talking, the delays are:
- Wait to be sure the caller finished speaking: 600 ms
- STT finishes the final text: 400 ms
- LLM sends its first token: 700 ms
- The system waits for the LLM to finish a full sentence before it sends text to TTS: 300 ms
- TTS makes the first audio: 350 ms
- Network to the caller: 150 ms

1. Add up the total delay before the caller hears the first sound.
2. Name the stage that is the biggest and the stage that is the easiest to cut.
3. Suggest two changes. For each, say how many ms it might save and what could go wrong.
Use a small table.

Now my team thinks about a realtime audio API. Audio streams in both directions, and the model starts to speak after about 450 ms in total.
1. Which stages of my old design disappear?
2. Give me two things I might lose, for example control, logging, or the choice of voice.
3. Give me one rule to decide between the two designs for a bank help line.
Answer in 8 lines or fewer.

I use a diffusion image pipeline. These tools exist: ControlNet (structure such as pose, depth, or edges), LoRA (a small trained add-on for a style or a character), inpainting (edit one masked area), and IP-Adapter (use a reference image as a guide).

For each job, pick the best tool, give one reason, and say what input I must have.
Job 1: Make 20 posters with the same cartoon girl in the same drawing style.
Job 2: Keep the exact pose of a person in my photo, but turn the photo into a painting.
Job 3: Change only the sky in a product photo and leave the product as it is.
Job 4: Make a new picture that has the colours and mood of one reference image, with no training.
If two tools could work, say so.

Take Job 2 and write the pipeline as a numbered list of nodes, as in ComfyUI. For each node, write its name, its input, and its output. Then give two reasons why a team should keep the pipeline as a saved graph file in version control.