1. The problem (20 seconds)
   “Typing receipt details into expense systems by hand is slow and error-prone. We built a system that reads a receipt image and extracts four fields: company, date, address, and total.”

2. Our approach (40 seconds)
   “We used Gemma, a model that can read images. We compared two versions:

• the original model, with no extra training (our baseline)
• a *fine-tuned model, which we taught using *LoRA, a lightweight method that trains a small add-on instead of the whole model

We trained on the SROIE dataset, which has receipt images with correct answers. We kept some receipts aside that the model never saw in training, so the test is fair.”

3. The pipeline
   “Receipt image goes into the model, which outputs JSON. We then compare that JSON to the correct answers and measure accuracy.”

4. Results (the most important part)
   Show your table with real numbers only:

Metric	Original	Fine-tuned
Company accuracy	[70]	[85]
Date accuracy	[60]	[75]
Address accuracy	[60]	[ 62]
Total accuracy	[70 ]	[ 80]
Exact-record accuracy	[ 50]	[55 ]
Invalid-JSON rate	[ 65]	[ 65]

If you don’t have fine-tuned results yet, say so: “We completed the baseline and data pipeline, and fine-tuning is.” Don’t invent numbers.
5. Errors and limitations
Mention things you actually saw, for example:

• confusing subtotal with total
• truncated addresses
• different date formats
• blurry receipts