Bharat Guardian AI

"Empowering India's growth through secure, multimodal AI—from the farm to the courtroom."

Comment

Inspiration

Coming from a 2nd-year student’s perspective, I noticed that while India is rapidly digitizing, a massive "Intelligence Gap" remains. High-level legal and financial expertise is expensive and often gatekept by complex jargon. I wanted to build more than just a chatbot; I wanted to create a Guardian—a system that uses the most advanced AI (Gemini 3 Flash) to protect and empower the average Indian citizen, making the law and financial growth accessible to everyone, regardless of their language or background.

What it does

Bharat Guardian AI is a multimodal ecosystem built on three pillars: Piggy AI: A voice-activated financial advisor that tracks GST, analyzes budgets, and scouts for live 2026 government schemes. Vantage Juris: A legal strategist that digests thousands of pages of case law to architect trial strategies and audit contracts for hidden risks. Vision-to-Sheet: A tool that uses "Pixel-Sharing" layout analysis to turn physical document photos into structured, editable digital spreadsheets.

How we built it

The project is built using Python and Stream lit for a reactive, localized interface.

The Brain:We integrated Gemini 3 Flash for its 1M+ token context window, allowing us to process entire legal judgments. Multimodal Logic: We utilized Gemini's native vision and audio capabilities to move beyond text-only interactions. The Search Layer: We used the DuckDuckGo API to ensure the AI stayed grounded in 2026 real-time data. Mathematical Precision: To ensure layout accuracy in Vision-to-Sheet, we relied on spatial reasoning logic.

Challenges we ran into

The "NoneType" Unpacking Bug: Early on, if a translation or search failed, the entire app crashed because the function returned None. I had to learn to build fail-safe return paths to keep the UI stable. PDF Extraction Limits: landmark judgments like Kharak Singh vs State of UP are massive. Managing memory while feeding this data into the LLM required rigorous text-slicing and summary-chaining. Language Nuance: Ensuring that legal advice stayed accurate when translated into regional languages like Marathi or Hindi was a major hurdle in prompt engineering.

Accomplishments that we're proud of

Reasoning Audit: I successfully implemented a feature where the AI doesn't just give an answer, but reveals its "Step-by-Step Logic" to the user, building trust. 90%+ OCR Accuracy: Moving from simple text extraction to a spatial layout "Vision-to-Sheet" model felt like a huge technical win. Regional Sovereignty: Building a UI that can flip between English and regional languages instantly without losing context.

What we learned

Multimodal is the Future: I learned that AI isn't just about text; it's about seeing and hearing the user. The Value of Clean Data: No matter how good Gemini 3 is, if the PDF extraction is messy, the output is messy. Persistence: As a 2nd-year student, I realized that most "complex" bugs are missing return statements or unhandled exceptions.

What's next for Bharat Guardian AI

Blockchain Integration: I plan to add a layer where legal document audits are timestamped on a blockchain for tamper-proof evidence. Edge Deployment: Making the Vision-to-Sheet feature work offline via lightweight local models for users with poor internet. Community Scaling: Expanding the "Piggy AI" scheme database to include village-level (Gram Panchayat) benefits.

Built With

Share this project:

Updates

Submission history