Loan Control Tower

• 

  AI scans every page, understands the structure, and locates the most important information.

• 

  Confidence scores are shown for every insight, supporting review and audit decisions.

• 

  Within minutes, the AI turns the document into a dashboard with critical dates and key terms.

• 

  From raw loan PDFs to structured insights, no matter the document format or length.

• 

  The AI explains each insight in simple, human-readable language.

Inspiration

I wanted to bring order to the wide range of loan and legal documents that exist today. There is no real standard in how these documents are written. Each agreement follows different structures, procedures, and outcomes, which makes them hard to follow and understand, especially after closing.

This problem is even clearer for someone without a legal background. I saw an opportunity to use AI to read, understand, and organize these documents in a consistent way, even when the documents themselves are not consistent.

What it does

Loan Control Tower ingests a loan agreement and turns it into an operational dashboard.

First, the document is uploaded and indexed. The AI reads the document to understand its structure and identify where key information is likely to appear. It then analyzes each relevant section and extracts structured data defined in a clear schema.

The results are presented in three operational views: Critical Dates (what’s coming next), Obligations (what must be done and when), and Risk Flags (default triggers and high-risk clauses).

Each item includes a confidence score and a plain-language explanation, and every insight is fully explainable with a direct quote and line reference from the original document.

 Target user

Loan Control Tower is built for post-closing loan operations teams at banks, lenders, and financial institutions, including loan operations, portfolio management, credit risk, and compliance teams.

It is also valuable for new team members onboarding into complex loan portfolios, where institutional knowledge is often fragmented or undocumented.

The product helps these users quickly understand obligations, monitor risk, and maintain compliance without re-reading hundreds of pages of legal text.

How we built it

This problem cannot be solved with rules alone, so I used an AI-first approach.

PDFs are ingested using an open-source framework called Docling, which converts them into a structured, line-anchored format for traceability.

I then use Gemini Flash 3.0 to analyze the document. AI is essential because loan agreements vary significantly in structure and language. The AI reads the document twice: first to understand and index its structure, and second to extract the specific data needed for the dashboard.

Even highly dense legal documents with hundreds of pages can be processed in minutes.

Challenges we ran into

The main challenge was handling the diversity of legal documents. There is no single way loan agreements present the same information.

I initially tried a rules-based approach using keywords and fixed logic, but it failed because the same concepts appear in very different ways across documents.

The solution was to rely on AI reasoning instead of rigid rules, allowing the system to first understand the document and then extract the data in a second pass.

Accomplishments that I’m proud of

The system can process a wide range of loan documents and still produce consistent, explainable outputs with a high level of confidence.

This required careful orchestration of the AI pipeline, many prompt iterations, and strong guardrails to avoid guessing. The result is a system that surfaces useful insights while remaining transparent and auditable.

What I learned

While building this project, I had to learn how syndicated loan agreements work, and I used AI extensively to do so.

In many ways, this app mirrors my own learning process: find the information, analyze it, structure it, and turn it into something understandable.

This highlights the real power of AI in complex domains, helping people achieve clarity even without deep prior expertise.

What’s next for Loan Control Tower

The next steps focus on making the product more production-ready and more valuable for users.

This includes improving data visualization, adding more interactive dashboards, introducing core application features like persistence, authentication, and security, and optimizing performance to reduce analysis time.

The long-term goal is to make Loan Control Tower a reliable post-closing system of record for loan teams.

Built With

• digitalocean
• docling
• fastapi
• gemini
• python