PDFSight · Owned product

Turning static PDFs into guided, fillable workflows.

PDFSight takes scanned, image-based, and non-interactive PDF forms through a complete workflow: identify the fields, collect the information conversationally, and return a filled document.

Live product

The problem

A form can look fillable without being usable.

A PDF can look like a form and still contain no usable fields. Scans, image PDFs, and flattened documents push the user back to printing, manual editing, or awkward re-entry.

The hard part is not generating text. The system has to interpret a visual document, turn that interpretation into structured fields, collect the right information, and produce a dependable file artifact.

Key decision

Use AI to understand the page. Use deterministic software to construct and fill the document.

PDFSight uses multimodal AI only where interpretation is required: reading the page and producing a structured field schema.

From there, conventional software takes over. The application creates and fills PDF fields, renders previews, and manages the resulting file through an expiring delivery path. The probabilistic step is bounded; the output pipeline is inspectable and testable.

Architecture

The system around the decision.

  1. 01ChatGPT / MCP guided UI
  2. 02Flask application API
  3. 03Vision model → field schema
  4. 04pypdf construction + filling
  5. 05Expiring artifact storage

What this demonstrates

What this work demonstrates.

PDFSight makes Amidship’s core thesis concrete: AI is valuable where interpretation is needed, but the product is the complete system around it. The user experience depends on workflow design, document processing, file lifecycle, interface behavior, testing, and infrastructure—not just field detection.

  • Product judgment around a narrow problem with a clear user outcome
  • Multimodal document understanding connected to deterministic software behavior
  • A deliberate boundary between probabilistic interpretation and reproducible output
  • ChatGPT / MCP tool and widget integration
  • API, preview, storage, expiration, and artifact-delivery engineering around the model

Inspectable evidence

Inspect it directly.

Open PDFSight (opens in a new tab)

Evidence boundary

PDFSight is an Amidship-owned product, not a client engagement. The public evidence demonstrates the product and its architecture; it does not claim customer adoption, accuracy, productivity gains, or commercial outcomes.