PAPERTRAIL: Turning Confusing Documents into Evidence-Backed Action Plans with AI
Understand the requirements. Know what to do. Never submit blindly.
Have you ever opened a scholarship notice, college circular, job application guideline, or government PDF and struggled to understand what you actually need to do?
The important requirements might be scattered across several pages. Deadlines may appear in different sections. Mandatory documents can be hidden inside tables, and one overlooked condition can cause unnecessary confusion.
A document can contain all the information you need and still be difficult to act on.
That is the problem I wanted to solve with PAPERTRAIL — The Clarity Engine, my project for the October 2026 MLH × DEV Writing Challenge.
The idea is simple: instead of merely summarizing a document, help people turn its instructions into a clear, interactive, evidence-backed action plan.
The problem with ordinary document chat
AI-powered PDF chat tools can summarize documents and answer questions, but users often need more than a summary.
They need to know:
- What exactly must I do?
- Which requirements are mandatory?
- What documents do I need?
- Which deadlines apply to which tasks?
- Where does the original document state this?
- Are any instructions potentially contradictory?
- What remains incomplete before I am ready to submit?
For important documents, a confident answer is not enough. Users also need a way to verify it.
That is where PAPERTRAIL takes a different approach.
Meet PROOFLINE: Every answer needs evidence
The central feature of PAPERTRAIL is PROOFLINE, an evidence-first approach to document intelligence.
When a requirement or answer is extracted, the system is designed to connect it to its supporting passage in the original document.
A citation can include:
- The source document.
- The relevant page and section.
- The original supporting text.
- A confidence indicator.
- A way to open the corresponding document location.
For example, if someone asks, “What is the application deadline?”, the goal is not simply to return a date. PAPERTRAIL should show the passage supporting that date and preserve its context.
This matters when a document contains multiple dates for different stages of a process.
The principle behind PROOFLINE is straightforward:
Don't just tell users what a document means. Help them verify why.
Citation validation can establish that a passage exists in the source, but it cannot guarantee that every interpretation is correct. Ambiguous or conflicting evidence still needs careful review.
How PAPERTRAIL works
1. Upload a document
PAPERTRAIL is designed to process supported formats including PDF, DOCX, TXT, and images.
The document becomes the starting point for an investigation into its structure, instructions, requirements, and important dates.
2. Extract requirements
Instead of stopping at a general summary, the requirement engine organizes information into actionable items.
These can include eligibility conditions, required documents, submission instructions, deadlines, exceptions, and important warnings.
Requirements are organized into stages such as preparation, eligibility verification, document collection, application completion, final review, and submission.
3. Explore the original evidence
The Smart Document Explorer connects extracted information to the source document.
Users can inspect the original wording rather than relying entirely on an AI-generated interpretation.
This is especially important when simplifying complicated instructions: plain language should make a requirement easier to understand without silently removing its conditions or exceptions.
4. Ask PAPERTRAIL AI
The document assistant is designed to answer questions using retrieved passages from the relevant document.
Examples include:
- Explain this paragraph in simple English.
- Which documents are required?
- What should I complete first?
- Translate this instruction into Hindi or Hinglish.
- Does this document contain potentially conflicting deadlines?
The system includes a deterministic grounded-answer mode, with optional integrations for Gemini and local Gemma inference through Ollama.
When an answer cannot be established from the available evidence, the appropriate response is to communicate that limitation rather than invent information.
5. Generate an interactive action plan
PAPERTRAIL converts requirements into a checklist that users can work through.
The action plan supports task completion, personal notes, priorities, reminders, and export options.
Progress is based on the user's actual checklist activity, not simply on the number of requirements extracted.
6. Run the “READY TO SUBMIT?” review
Before finishing a process, users can review incomplete tasks, unresolved questions, unverified requirements, and possible contradictions.
The purpose is to help people prepare more carefully—not to claim that an official application is eligible, approved, or successfully submitted.
More than a document chatbot
PAPERTRAIL also includes several supporting capabilities:
Contradiction detection: Highlights potentially conflicting passages so users can inspect both sides.
Deadline intelligence: Organizes dates with their context and source references.
Document comparison: Helps identify changed requirements, deadlines, fees, and instructions between document versions.
Requirement graph: Visualizes relationships between stages and tasks.
Plain-language translation: Provides simpler explanations in English, Hindi, and Hinglish.
Multi-document workspaces: Designed to preserve source references when working with related documents.
Accessibility controls: Includes features such as text scaling, high contrast, reduced motion support, and light and dark themes.
The aim is to make complex instructions easier to work with, including for people who are not technically experienced.
The technology behind PAPERTRAIL
The application uses a modular full-stack architecture.
Frontend
- React 19
- TypeScript
- Vite
- Tailwind CSS
Backend
- Python
- FastAPI
- Pydantic
Document processing
- PyMuPDF for PDF processing
- python-docx for Word documents
- Pillow for image processing
AI and retrieval
- A deterministic grounded-answer engine
- Retrieval-Augmented Generation (RAG)
- Optional Gemini integration
- Optional local Gemma inference through Ollama
Storage and testing
- Local JSON-based workspace persistence
- Automated backend tests
- A synthetic document package with ground-truth evaluation data
The deterministic mode is intended to provide useful document processing and evidence-based responses without requiring a paid AI API. Optional model integrations depend on their respective configuration and availability.
Measuring the results
PAPERTRAIL includes a synthetic evaluation package called The Application Clarity Challenge.
According to the project's benchmark report, the evaluation produced the following results:
- Text extraction: 3 out of 3 pages processed successfully
- Requirement extraction: 100% precision and 100% recall across 17 benchmark items
- Date extraction: 100% accuracy across 4 benchmark dates
- Citation verification: 100% of evaluated quotes were verifiable
- Contradiction detection: 100% recall on the benchmark's tested deadline and fee conflicts
- Retrieved passage relevance: 100% on the benchmark
- Reported pipeline latency: 135 ms
The project's test report also records eight passing automated tests.
These results come from a small, synthetic evaluation dataset. They are useful for checking the implementation against known examples, but they do not establish equivalent performance across real-world documents, scanned PDFs, different languages, or complex policies.
That distinction is important. Reliable AI products should communicate what has actually been tested, not turn a successful demonstration into an unsupported universal claim.
Privacy and responsible document handling
Documents can contain personal and sensitive information, so privacy is an important design consideration.
PAPERTRAIL supports a local-first workflow and optional local model inference. Local inference can reduce the need to send document contents to an external AI provider, provided the rest of the application also keeps processing local.
Optional cloud inference requires appropriate configuration and careful handling of document data.
The project is intended to help people understand and prepare their applications. It does not replace official guidance, guarantee eligibility, or submit applications automatically.
What building PAPERTRAIL taught me
One of the biggest lessons from this project is that AI functionality alone does not make a useful product.
A document assistant needs more than a model that generates fluent answers. It needs reliable parsing, source tracking, meaningful retrieval, honest uncertainty handling, useful task management, and tests that measure its behavior.
I also learned how important it is to distinguish between extracting information and interpreting it.
A date can be extracted correctly but assigned the wrong meaning. A passage can be cited correctly but fail to support the conclusion drawn from it. A checklist can look complete while the user still needs to verify a condition.
These are reasons to design the entire workflow around evidence, not just the model.
There is still room to expand the evaluation dataset, test more diverse documents, improve OCR and multilingual handling, and validate performance on real-world examples.
Try PAPERTRAIL
🌐 Live demo: https://raghuraj72.github.io/PAPERTRAIL/
💻 GitHub repository: https://github.com/raghuraj72/PAPERTRAIL
Explore the demonstration and repository to see how the document workspace, requirement extraction, action planning, and evidence-first approach fit together.
Some AI and document-processing features may require the backend or a configured model to be running locally. Check the repository's setup instructions for the complete experience.
Final thought
A complicated document should not leave someone wondering what to do next.
People deserve to understand the requirements, identify the evidence, organize their tasks, and know what still needs attention.
PAPERTRAIL is my attempt to make that process clearer, more transparent, and more actionable.
PAPERTRAIL — Turn confusion into clarity.
Built for the October 2026 MLH × DEV Writing Challenge.
Top comments (0)