Achieving 90% Accuracy Using GPT-5.4

Hello Junaidh Ahamed,

Greetings! Thanks for raising this question in the Q&A forum.

Short answer: 90% or better is realistic for signature and handwriting detection, but I would not promise it for stamps and overlapping marks with GPT-5.4 alone. Nobody can guarantee a number in advance, because accuracy depends on your own documents, so the right way to answer "is it feasible" is to measure it on a labeled sample of your own files. A multimodal model like GPT-5.4 is good at understanding a page and answering "is there a signature, a seal or handwriting here", but it is weaker at precise localization (exact bounding boxes), at very small marks, and at giving the same answer every time on low quality scans. A hybrid pipeline usually performs better and is easier to tune.

Here is the approach I would recommend:

Build a labeled test set first. Collect 200 to 500 representative pages that include clean scans, low quality scans, overlapping signatures and stamps, and faint handwriting. Label each page for signature, stamp or seal, and handwriting (and bounding boxes if you need locations). Measure precision and recall per class, not only overall accuracy, because a page without any signature is easy to get right and can hide poor recall.

Use Azure Document Intelligence for handwriting and layout. The Read and Layout models return lines and words with their positions, and they report handwritten style when it is detected, so you get handwriting detection with coordinates without any extra model. Enable the style font add-on and the high resolution add-on for small or faint text. Document Intelligence also returns paragraphs, selectionMarks and other structure you can use to know where a signature block should be.

Use a custom model for signatures. Document Intelligence custom template or custom neural models let you label signature fields and return whether a signature is present in that field. This works well when documents share a layout (for example, a signature line at the bottom of a form). Note that the prebuilt models do not identify whose signature or which company a stamp belongs to, they only extract what is visible.

Consider Azure AI Content Understanding for a single schema driven call. Content Understanding lets you define a schema (for example fields such as has_signature, has_company_stamp and has_handwriting) and uses a multimodal model with confidence and grounding information. It is a good fit when you want structured output per document and do not want to manage several models. Check the current preview or GA status and region availability before you commit to it, because this service is changing quickly.

Use GPT-5.4 for the hard cases, with a strict prompt. Send page images at high detail, ask for JSON only with a fixed schema (signature_present, stamp_present, handwriting_present, plus a short reason and a confidence from 0 to 1), use structured outputs, and set temperature low. Giving two or three labeled example images in the prompt (few shot) usually helps with stamps and overlapping marks. Cropping the region of interest (for example the bottom third of the page, or the region around a Document Intelligence handwritten span) and sending that crop at higher resolution improves small mark detection.

Add a specialized vision model if stamps are the weak spot. For precise bounding boxes of stamps, seals and signatures, a trained object detector (for example a YOLO based model trained on your own stamps, using Azure AI Custom Vision or Azure Machine Learning) is often more accurate and cheaper per page than an LLM. A common design is: detector first, then GPT-5.4 only for low confidence results.

Improve input quality before detection. Render PDFs at 200 to 300 DPI, deskew, denoise, and normalize contrast. Avoid heavy compression. For overlapping stamps and signatures, colour matters, so keep colour images and do not convert to grayscale too early, since many stamps are red or blue and can be separated by colour channel.

Route by confidence and keep a human in the loop. Accept results with high confidence automatically, and send low confidence or disagreement between two methods (for example Document Intelligence and GPT-5.4) to a person for review. Feed the reviewed pages back into your test set and training data. This is usually how teams reach and keep 90% or higher in production.

Check cost, latency and quotas. Image tokens for GPT-5.4 on every page can be costly at volume, so a layered pipeline (cheap model or detector first, GPT-5.4 for uncertain cases) is normally the more economical choice. Confirm the model's availability, regions and rate limits in your Foundry deployment.

If this answer helps you kindly accept the answer which will help others who have similar questions.

Best Regards,
Jerald Felix.