ShieldDoc AI

ShieldDoc AI removes personal data first, then helps you understand documents—what happened, what you owe, and what to do next—safely.

Comment

Inspiration

Every day, people paste sensitive documents—medical bills, insurance reports, contracts—into AI tools to understand them. Most don’t realize their personal data is being exposed in the process. We wanted to solve that gap: how do you get the power of AI without sacrificing privacy?

What it does

ShieldDoc AI lets users upload any document, automatically removes sensitive information, and then allows them to safely ask questions.

It goes beyond redaction by:

extracting key information (costs, diagnosis, follow-ups) highlighting important notices explaining documents in plain English suggesting next steps answering questions with source grounding providing contextual explanations for “why” questions How we built it Frontend: HTML, CSS, JavaScript Backend: Python (Flask) PDF Processing: PyPDF2 Redaction Pipeline: Regex-based removal (emails, SSNs, phone numbers, dates) AI-assisted redaction (names, providers, IDs) AI System: Gemini API integration Strict grounded Q&A mode (document-only answers) Contextual reasoning mode (for “why/necessary” questions) Reliability: Local regex fallback extraction ensures key data is always available UI Features: Dynamic suggested questions Evidence highlighting Real-time redaction summary Full state reset system Challenges we ran into AI response truncation breaking JSON parsing Separating strict answers vs reasoning answers Preventing hallucinated insights Ensuring accurate redaction with both regex + AI Managing UI state without stale data Accomplishments that we're proud of Built a full privacy-first AI pipeline in 24 hours Successfully combined deterministic logic with AI for reliability Created dual-mode AI (factual + reasoning) Delivered a polished, real-world usable product—not just a prototype What we learned Privacy must be built into the pipeline, not added later AI requires strong guardrails (grounding, structure, fallback) Small UI details significantly impact perceived quality Combining rule-based systems with AI improves reliability What’s next for ShieldDoc AI Fully local redaction pipeline for stronger privacy guarantees Support for more document types (legal, financial, academic) Smarter follow-up question generation Improved explanation quality and personalization Key Idea

ShieldDoc AI separates privacy from intelligence: we clean the data first, then let AI do the thinking.

Built With

Share this project:

Updates