Hitachi Innovation

Inspiration

The idea for this project came from something I kept noticing. People and teams waste hours sorting through documents just to figure out what is confidential, safe to share, or needs review. It is repetitive and time-consuming, and one small mistake can cause major issues. I have always been drawn to building things that make people’s work easier and smarter, and this felt like a problem that AI could genuinely solve if used the right way. I wanted to create something that not only analyzes content but understands the context behind it, something people could actually trust to help them.

What it does

The system reads through multi-page documents that include text, images, and even scanned forms, and classifies them into categories such as Public, Confidential, Highly Sensitive, or Unsafe. It shows the reasoning too by pointing to the exact page or image that led to the classification. It works in both interactive and batch modes, provides real-time updates, and makes it easy to understand what is happening behind the scenes. Instead of just labeling content, it explains why, and that makes it transparent and practical.

How we built it

We built it using Python and FastAPI for the backend, React and Chart.js for the dashboard, and integrated AI models using OpenAI APIs. We also implemented a Human-in-the-Loop system where experts can correct or confirm classifications, and the AI learns from that feedback over time. The goal was to make something explainable, secure, and constantly improving. This system was built to be modular, explainable, and reliable. Every part of it was designed to allow the backend, AI models, and interface to evolve without breaking the others. We wanted it to be easy to maintain, scalable, and secure, while keeping everything transparent and auditable. The system is divided into smaller, independent modules for classification, preprocessing, and evidence generation. Each classification comes with clear reasoning and citations, so anyone can see why the AI made that decision. This makes the system transparent and easy to debug or expand. We built a feedback system where subject matter experts can review AI decisions, approve or correct them, and help the system improve over time. Every review is logged and contributes to better accuracy and confidence scoring in future classifications. This makes the AI more dependable without removing human oversight. We also implemented AES-256 encryption for stored files, SSL/TLS for all network communication. Each document and classification result is logged in an auditable database for traceability and compliance verification. No sensitive data leaves the secure cloud environment.

Challenges we ran into

The hardest part was getting the AI to justify its decisions. It is easy to classify, but it is hard to explain. Teaching it to point out specific evidence while still staying fast took a lot of experimentation. Handling both text and images together also added complexity since the system had to process multiple formats and merge the logic into one output. Another major challenge was finding a cost-optimal model that could still perform with high accuracy at scale. We had to balance performance with efficiency so the system could handle batch processing without becoming expensive to run. Making sure the system detected the right number of pages and images for each file was also essential for consistent validation.

On top of that, integrating reinforcement learning for the Human-in-the-Loop verification took time to get right. We wanted the AI to genuinely learn from human feedback and adjust its confidence over time instead of just storing corrections. Additionally, we wanted the system to be able to learn from previous company interactions and decisions so new users could benefit from that knowledge. This allows the AI to suggest actions for similar cases in the future and helps companies maintain consistent compliance decisions over time.

Accomplishments that we're proud of

The biggest accomplishment is that our AI does not just give answers, it gives proof. Every classification comes with reasoning, citations, and traceability. It feels like something that people could actually use in a real-world setting rather than just a demo. We are proud that it bridges the gap between automation and human understanding. It is explainable, safe, and works with people rather than against them.

What we learned

We learned that trust matters just as much as accuracy. People want to see how AI thinks before they accept what it says. We also realized how important user experience is. A clear interface and understandable results change everything about how someone feels when using the system. This project helped us understand how to balance technical power with usability and communication.

What's next for Hitachi Innovation

Next, we plan to make the system even smarter. We want to add automatic redaction for PII, multilingual support, and deeper compliance policy understanding. We also plan to visualize how human feedback improves the AI over time to create a transparent feedback loop. In the long term, we want this to become a reliable and practical tool for organizations that want to manage documents safely, efficiently, and confidently. It is about taking something overwhelming and turning it into clarity that people can rely on.