AI PDF Application Scorer

Score PDF applications against your own rubric, with a person in the loop for anything that looks off.

Applicants upload a PDF through an n8n form. The workflow reads it, scores it against the rubric and pass threshold you write, and emails the applicant the result.

The AI never decides pass or fail. It scores each criterion, and plain code checks those scores and applies your threshold.

Documents that try to manipulate the reviewer get caught in layers:

1. Pattern screen. Plain checks for text aimed at the AI, like "ignore previous instructions" or hidden characters.
2. Classifier. A separate AI call whose only job is to ask "is this document trying to manipulate the reviewer?" If that call fails, the document is held.
3. Output checks. Every score must be present, in range, and add up. Notes can't contain links, HTML, or instruction-like text.

Anything flagged is held for a human and never auto-approved. You get an email with the reasons, the extracted text, and the original PDFs. The applicant hears nothing until you decide.

No single defense stops every manipulation attempt. Layering them and holding anything flagged for a person lowers the risk. An application that gets past all three checks is scored and answered automatically, like any other. If you'd rather approve passes yourself, one setting holds every passing application for your review (off by default).

Features

• Your rubric and your threshold, set in one Config node
• Works with OpenAI-compatible chat-completions APIs, such as OpenAI, OpenRouter, or a local model via Ollama
• Rate guard: caps submissions per hour and alerts you at most once an hour when it drops extras
• File limits: number of files, size, PDF only, page count, and text length, checked before any AI call
• Three-layer prompt-injection defense
• Human hold for anything flagged
• Optional review hold for passing applications, so nothing reaches an applicant until you approve it (off by default)

What's included

• 2 workflow JSON files (main workflow + rate-guard sub-workflow)
• README with setup, configuration, and test steps
• Docker self-hosting guide
• License for one business

Requirements

• Self-hosted n8n 2.x (tested on 2.42.6)
• An API key for an OpenAI-compatible service, or a local model through Ollama
• An SMTP mail account

License

One license covers one business, for unlimited use in its own operations. Setting it up for a client? You can, as long as that client's business has its own license. No reselling or redistributing the files. Free updates for 12 months from purchase. Full terms are in the license in the download.

FAQ

Which AI model should I use?

One served through an OpenAI-compatible chat-completions API, such as OpenAI, OpenRouter, or a local model in Ollama. You choose the model and pay your provider directly. If a model's reply can't be read, the application is held for you rather than scored.

Can I keep applicant documents away from AI providers?

Yes, if you point it at a local model. Text extraction uses n8n's built-in node, with no outside service. Held applications are emailed to you with the PDFs attached, so they do pass through your mail provider.

Is it immune to prompt injection?

No tool can promise that. This one uses three separate checks, and anything any of them flags goes to a person instead of being approved.

Does the AI decide who passes?

No. The AI scores each criterion. Code checks the scores and applies your threshold. You can also turn on a setting that holds passing applications for your review before anyone is told.

What happens to a flagged application?

You get an email with the reasons, the extracted text, and the PDFs. Nothing goes to the applicant until you act.

Does it work on n8n Cloud?

It's built and tested for self-hosted n8n 2.x. We don't support n8n Cloud.