Intelli-Credit

An autonomous AI Credit Officer for explainable corporate underwriting and capital impact simulation.

Comment

Inspiration

Corporate underwriting at Tier-1 banks is a trillion-dollar bottleneck. It relies on a slow, manual synthesis of the "Five C's of Credit" and rigid rule-based engines. While modern Machine Learning can accurately predict default probabilities, these standalone models are "black boxes" that credit committees fundamentally distrust. We realized that to truly fix underwriting, we couldn't just build a better calculator—we needed to simulate the actual cognitive workflow and debate of a real credit committee. We were inspired to build an autonomous system that marries mathematically rigorous, explainable ML with the qualitative nuance of human underwriters.

What it does

Intelli-Credit is an autonomous AI Credit Officer. Instead of just outputting a simple credit score, it evaluates a corporate borrower exactly how a bank would:

  1. Autonomous Due Diligence: It ingests both structured financial data and unstructured statutory filings, conducting web-scale research for ESG sentiment and litigation risks.
  2. Explainable Risk Scoring: It predicts the Probability of Default (PD) and recommends credit limits, explicitly extracting the top 5 contributing risk drivers so the decision is 100% transparent.
  3. Capital Impact Simulation: It evaluates the portfolio impact using Risk-Adjusted Return on Capital (RAROC) stress tests.
  4. Agentic Debate: It utilizes a multi-agent swarm to simulate a credit committee debating the qualitative risks.
  5. CAM Generation: Finally, it synthesizes all this data into a downloadable, professionally structured 8-section Credit Appraisal Memo (CAM) in PDF format.

How we built it

Intelli-Credit is built on a decoupled, high-performance architecture:

  • The Orchestrator (Backend): We used Python 3.11 and FastAPI for async-native performance. The deterministic risk models were built using Scikit-Learn and XGBoost, wrapped in SHAP (SHapley Additive exPlanations) to crack open the algorithmic black box.
  • The Trust & Swarm Layer: To evaluate qualitative risk, we integrated the core concepts of MiroFish (a swarm intelligence prediction engine). Using CAMEL-AI and Mem0 (GraphRAG), we spawn specialized AI agents (e.g., a bearish Risk Manager and a strict Compliance Officer) that query a dynamic knowledge graph of the borrower and "debate" the risk until a consensus emerges.
  • The Financial Physics Engine: We built a custom simulation module to calculate capital impact. The Expected Loss (EL) is derived dynamically from our trained XGBoost ensemble:

$$ \text{EL} = \text{PD} \times \text{LGD} \times \text{EAD} $$

This is then fed into our portfolio stress-tester:

$$ \text{RAROC} = \frac{\text{Expected Revenue} - \text{Expected Loss (EL)} - \text{Operating Costs}}{\text{Economic Capital}} $$

  • The Command Center (Frontend): A deeply interactive B2B SaaS portal built with Next.js 16, React 19, and TailwindCSS 4. We implemented real-time cognitive flow animations and a visual node-based decision studio powered by @xyflow/react.

Challenges we ran into

  • Latency vs. Intelligence: Running a multi-agent debate using CAMEL-AI can easily result in API token explosions and severe latency. We had to aggressively optimize our prompts and utilize hierarchical GraphRAG summaries to achieve near-real-time SaaS responsiveness.
  • The Unstructured Data Swamp: Extracting clean, tabular data from scanned, inconsistent Indian statutory filings (like GSTR-3B vs. 2A reconciliations) was brutal. Building a resilient ingestion pipeline that elegantly falls back from regex extraction to OCR (pytesseract + pdfplumber) was a massive engineering sprint.
  • Fusing Two Worlds: Ensuring the non-deterministic LLM swarm didn't hallucinate over the hard, deterministic math produced by the XGBoost models required building a strict logic rules engine to act as the final arbiter before the PDF CAM generation.

Accomplishments that we're proud of

  • True Explainability: We successfully solved the "black box" problem. Seeing the UI dynamically render the exact SHAP values explaining why a limit was reduced by 20% is a game-changer for enterprise trust.
  • MiroFish Swarm Integration: Taking the theoretical concept of multi-agent swarm prediction and successfully applying it to a rigorous financial underwriting workflow.
  • The User Experience: We are incredibly proud of the frontend. Translating complex agentic reasoning and backend ML orchestration into a beautiful, intuitive node-based UI makes the platform feel like magic.

What we learned

Building Intelli-Credit completely reshaped our understanding of AI in the financial sector. We learned how to architect complex systems where deterministic ML algorithms and generative multi-agent swarms operate symbiotically... the swarm provides the context, and the math provides the boundaries. We also deepened our full-stack engineering skills, particularly in managing complex, asynchronous state across a Next.js frontend while streaming live orchestration updates from a FastAPI backend. Most importantly, we learned that in the enterprise world, an AI's explainability is infinitely more valuable than marginal improvements in its raw statistical accuracy.

What's next for Intelli-Credit

We aren't just building a SaaS application; we are building the intelligence infrastructure for the next generation of autonomous lending. What's next is expanding our GraphRAG capabilities to ingest deeper alternative data sources (supply chain networks, satellite data for ESG) and allowing risk managers to deploy custom Python-based overriding nodes directly within the browser studio. Ultimately, Intelli-Credit paves the way for a future where capital is allocated safely, transparently, and at the speed of software.

Built With

Share this project:

Updates