Inspiration
The world of credit is stuck in the 1990s. Traders and Credit Officers are still manually skimming 300-page PDF credit agreements at 2 AM, relying on "Ctrl+F" and caffeine to find billion-dollar risks. We realized that while high-frequency trading is algorithmic, private credit - a $1.5 trillion market - is still painfully analog and most of the times involves the lovely green IBM Terminal that will never die.
We asked ourselves: What if a credit agreement wasn't a document you read, but a dashboard you drove?
We wanted to build the "Formula 1" cockpit for credit: stripping away the legalese to present only mission-critical data, allowing traders to move from manual reading to algorithmic execution.
What it does
Ocorm.com is a document ingestion engine that transforms chaotic deal rooms into a single, intentionally designed "One Long Page" dashboard. It aggregates every disparate piece of data and sorts it dynamically based on exactly who is looking at the screen.
Deep Detective Work & Proprietary Scoring: It doesn't just read text; it hunts. Unlike a human, it doesn't get tired at Page 1,000, and it doesn't lose focus just because one document was reviewed by a sleep-deprived intern and the other by a senior partner, late to a meeting. The engine digs out risks buried deep in footnotes, re-evaluates stale appraisals against real-time market data, and synthesizes everything into a Proprietary Risk Score. This allows you to spot a "B+" credit disguised as a "C-" mess (or vice versa) in seconds.
Bespoke Comparison (Your Docs vs. Theirs): We are not trying to standardize the market. Every fund is unique. The engine compares the incoming credit agreement against your internal standard playbook, your historical yields, and your specific risk appetite. It highlights exactly where this deal deviates from what you usually sign, rather than forcing you into a generic "market standard" that doesn't fit your strategy.
The "Karen" Automation: We realized that trying to change internal processes is impossible because Karen from Risk (empowered by Bob from Legal) will simply never let go of her 20-page internal Excel checklist. Instead of fighting it, we embrace it and automate it. The dashboard pre-fills Karen’s rigid spreadsheet perfectly. The trader can shove it to the bottom of the page and ignore it, but Karen stays happy, and the deal gets approved.
The "Tired Joe" Effect: The dirty secret of finance is that Tired Joes kill the most lucrative deals. A deal with massive documentation usually means it’s highly collateralized and safe - but to Joe, who played video games all night, it just looks like work. Historically, Joe will find some flimsy excuse to reject a complex deal simply because he doesn't want to spend his weekend cross-referencing numbers in a legacy IBM terminal. By serving the deal on a silver platter, we ensure Joe doesn't kill a gold mine just because he wants to go home.
How we built it
We moved beyond simple "Chat with PDF" wrappers to design a Multi-Agent Orchestration Engine rooted in a paranoia-first architecture. You cannot simply ask ChatGPT to summarize a credit agreement - it will hallucinate. To solve this, our framework deploys a swarm of 5 specialized agents to populate the data:
- The 5-Agent Swarm Legal Agent ("Lawyer in a Box"): Reads docs to validate collateral (UCC-1s) and compares every clause against your specific internal playbook using case-law trained NLP.
Quant Agent (Financials & Math): Ingests borrower financials to calculate core loan mechanics (Yields, DSCR) and predict future portfolio performance.
Appraisal Agent (Valuation): Automates valuation by extracting physical specs from third-party reports (CBRE) and benchmarking them against 10M+ real-time market comps.
Risk Agent (The Guardian): Acts as the system's internal auditor, cross-checking every agent's output for mistakes and hallucinations to ensure you never commit to illegal or non-compliant terms.
ESG Agent (Sustainability): Checks the loan against Green Loan Principles and exclusion lists, screening the borrower and collateral to ensure the deal itself meets your sustainability mandates.
Challenges we ran into
The "Table Problem" (Geometry vs. Text) Financial data isn't just text; it’s geometry. We found that standard OCR completely fails on complex, nested financial tables. The agents struggled to understand that a "Q4 EBITDA" header applied to a number three rows down and across a page break, turning critical data into unstructured noise.
The Data Gap & Hallucination Risk We faced a hard reality: we cannot rely on off-the-shelf LLMs for institutional finance. They lack the specific context for high-stakes credit and often hallucinate "standard" clauses where none exist.
The "Karen vs. Joe" UX Paradox We were stuck between two diametrically opposed users. Karen from Risk demands to see every single audit log and footnote, while Tired Joe (the Trader) wants to see absolutely nothing but the final yield and red warning. Trying to design a single dashboard that didn't terrify Joe with density while still satisfying Karen's need for compliance was our most difficult design conflict.
Accomplishments that we're proud of
Universal Market Validation: We conducted deep market analysis and spoke with real professionals across the industry. We haven't found a single person involved in loans - from overworked analysts to senior partners - who didn't immediately love the simplicity.
Cracking the "Karen vs. Joe" Paradox: We successfully designed a solution for the industry's biggest friction point: how to give Risk officers (Karen) the granular depth they legally require without paralyzing Traders (Joe) with data they hate. Our wireframes proved that a single interface can serve both masters.
The "One Long Page" Breakthrough: We demonstrated that the cure for "Deal Room Fatigue" isn't more tabs or complex dashboards- it's a linear, scrolling narrative. The feedback confirmed that our "intern-simple" design is exactly what the market is starving for.
What we learned
Trust is a UI Feature: Traders don't trust black boxes. We learned that the most critical feature isn't the AI model itself; it is the Source Link. If the user can't click "View Appraisal" and see the original paragraph immediately, they simply won't believe the risk score.
Adjectives don't matter; Data does: "Good condition" means nothing in finance. "NVIDIA H100 Cluster, 36 Months Remaining Life" means everything. We learned that for an AI agent to be useful, it must be ruthless about stripping out fluff and extracting hard specs.
The "Alt-Tab Tax" is Real: The biggest friction in deal-making isn't the difficulty of the math; it's the cognitive load of switching between 12 different PDFs, 5 different system and trying to explain Karen's spreadsheet to your syndicate partner. Centralizing everything into One Long Page creates a psychological "ease of use" that actually leads to faster, sharper risk decisions. Even though its long and has a lot of data points that can take days to fully review. Caters to all -clear summary, available details.
Simplicity is the Ultimate Scale: We initially thought we needed complex charts. We learned that "Intern-Simple" is actually the highest form of sophistication. If you can explain a $100M deal in a simple list that a tired analyst can read on a phone, you have solved the hardest problem in the room. Lawyers are still there to make sure no one lost their minds in the process. We just switched them from analyzers to verifiers and simplified their life too.
What's next for Simplest All-in-One AI Loan Review & Diligence Suite
Strong partner. We currently lack the capital and proprietary data agreements required to train specialized models. Our immediate next step is securing the funding and legal partnerships to access private credit datasets, which are the only way to move beyond generic LLMs and build a truly "hallucination-proof" engine.
Closing the Data Gap: We will move from general LLMs to fine-tuning our own small language models (SLMs) specifically on proprietary credit agreements. This will eliminate hallucinations and teach the agents the nuance of "Permitted Liens" that off-the-shelf models miss.
Direct ERP Integration: Moving beyond "Ingesting PDFs" to "Ingesting the Borrower." We aim to connect directly to a borrower’s QuickBooks or ERP to automate the Quant Agent's financial analysis in real-time.
Built With
- javascript
- llama
- next.js
- php
Log in or sign up for Devpost to join the conversation.