Inspiration Financial risk analysis often requires hours, sometimes days, of manually gathering market data, investigating news, cross-referencing sources and compiling institutional reports. Critical warning signs can easily get buried across fragmented information sources.
We wanted to address this problem by building ShadowRisk AI: an automated forensic intelligence engine designed to help human analysts uncover what traditional financial metrics alone cannot reveal. Our goal was simple: transform scattered information into structured, evidence-based risk intelligence without sacrificing analytical depth or reliability. What it does ShadowRisk AI is an automated forensic credit and risk intelligence platform that combines real-time financial market data with AI-powered web intelligence to generate institutional-grade risk assessments in seconds.
Given a company name and ticker symbol, it retrieves financial metrics through Yahoo Finance and investigates four critical risk pillars: supply chain, legal, labor and regulatory exposure.
Rather than simply summarizing search results, ShadowRisk AI synthesizes the findings into structured credit reports featuring evidence tables, source citations, risk assessments and institutional recommendations.
Its goal is to help analysts identify hidden vulnerabilities, understand potential exposure and make more informed decisions. How we built it We built ShadowRisk AI using a Python-based FastAPI backend and a web-based frontend.
- Financial Intelligence: Yahoo Finance integration retrieves real-time market metrics, including market capitalization, P/E ratio, beta and historical price ranges.
- Multi-Pillar Research: Tavily powers parallel intelligence gathering across supply chain, legal, labor and regulatory risk categories, with URL deduplication and content filtering.
- AI Report Generation: An LLM processes the collected evidence and generates structured institutional-grade credit assessments.
- Deterministic Risk Scoring: The final ShadowRisk Score is calculated server-side using a strict weighted algorithm rather than relying on the LLM for arithmetic.
- Resilient Generation Pipeline: A 16,000-token output budget, custom brace-matching JSON parser and automated self-repair mechanism help prevent incomplete or malformed reports.
- Provider-Agnostic Architecture: Automatic fallback routing allows the application to switch LLM providers when the primary provider encounters rate limits.
- Report Export: Analysts can export their completed assessments as clean, multi-page PDF documents. We also prioritized a local-first deployment approach, allowing users to run the application on their own machines while configuring their API providers independently. Challenges we ran into One of our biggest challenges was generating lengthy, structured AI reports reliably.
Institutional-grade assessments can span three to four pages, pushing models beyond their reliable output limits. This frequently creates truncated responses and broken JSON structures, making reports impossible to process.
To address this, we implemented a custom brace-matching parser and an automated self-repair loop that detects incomplete outputs and feeds partial responses back to the model for correction.
Another major challenge was preventing AI-generated numerical inaccuracies. Instead of allowing the LLM to calculate the final risk score, we moved the entire scoring process to a deterministic backend algorithm.
We also addressed provider reliability through automatic fallback routing, ensuring that rate limits from one provider would not unnecessarily interrupt the analysis workflow. Accomplishments that we're proud of We are particularly proud of building a complete forensic risk intelligence workflow that brings financial data, multi-source research and AI-generated institutional reporting into one application.
Some of our key accomplishments include:
- Automating research across four distinct risk pillars.
- Developing a resilient long-form report generation pipeline.
- Implementing deterministic risk scoring to prevent arithmetic hallucinations.
- Building automatic LLM provider fallback routing.
- Delivering structured, source-backed reports with PDF export.
- Creating a local-first, provider-agnostic architecture. Most importantly, we built a tool designed to assist human analysts rather than replace their judgment, giving them a faster way to investigate complex risks and organize their findings. What we learned Building ShadowRisk AI taught us that developing reliable AI applications involves much more than connecting an LLM to an API.
We learned the importance of separating deterministic calculations from probabilistic model outputs, designing robust parsing and recovery mechanisms, managing external API dependencies and validating information across multiple sources.
We also gained valuable experience integrating financial data providers, search infrastructure and LLM services into a unified application.
One of our biggest takeaways was that reliability, traceability and evidence quality are just as important as model intelligence, especially when building tools intended for institutional decision-making. What's next for ShadowRisk AI We envision ShadowRisk AI evolving into a more comprehensive enterprise risk intelligence platform.
Our future priorities include:
- Expanding coverage to additional industries, financial instruments and risk categories.
- Improving source verification, evidence traceability and confidence assessment.
- Introducing historical risk tracking to identify emerging patterns and changes in corporate exposure.
- Expanding deployment options for enterprise environments and private infrastructure.
- Further optimizing inference performance and multi-provider orchestration. Our long-term vision is to make institutional-grade forensic intelligence more accessible, efficient and reliable, helping analysts see beyond financial metrics and uncover risks before they become major problems.
Log in or sign up for Devpost to join the conversation.