Inspiration
Security teams face a prioritization problem: thousands of vulnerabilities are disclosed each year, but only a small fraction are actively exploited. Traditional approaches relying on CVSS severity, EPSS, and public exploit signals do not always identify which vulnerabilities will become real-world threats.
We built PatchRadar to address this gap using machine learning and IBM Granite, helping organizations prioritize vulnerabilities based on predicted exploitation risk rather than severity alone.
What it does
PatchRadar is an AI-powered vulnerability intelligence platform that predicts and ranks the 500 CVEs most likely to be added to CISA's Known Exploited Vulnerabilities (KEV) catalog.
The platform provides three core capabilities:
- Predictive Risk Ranking: An ensemble machine learning model identifies high-priority CVEs using vulnerability metadata, exploitation indicators, and IBM Granite-derived text features.
- Interactive Threat Analytics: A React dashboard visualizes vulnerability trends, historical exploitation patterns, and the model's top 500 predictions.
- AI Threat Analyst: An IBM watsonx.ai-powered assistant translates natural-language questions into predefined DuckDB queries and delivers evidence-backed summaries.
The business value is straightforward: help security teams identify higher-risk vulnerabilities earlier and allocate limited remediation resources more effectively.
How we built it
Data Engineering — Python, Pandas, DuckDB
Processed 64,120 historical CVE records, engineered temporal and exploitation-related features, and established publication-time feature constraints to prevent future-information leakage.
Machine Learning — scikit-learn, LightGBM, IBM Granite
Developed a 50/50 ranking ensemble combining balanced logistic regression and LightGBM, with the latter averaged across 30 seeds. Integrated IBM Granite-derived vulnerability text scores and engineered CVE metadata features to improve risk prioritization.
Evaluated the model using chronological backtesting, repeated grouped cross-validation, and rolling-window testing against a conventional exploit-signal and EPSS baseline.
AI Integration — IBM watsonx.ai, FastAPI, DuckDB
Built a natural-language analytics service that maps questions to predefined SQL queries, retrieves structured vulnerability data, and uses watsonx.ai to summarize the results. Implemented deterministic fallbacks and rate limiting for reliability and API cost control.
Frontend — React, TypeScript, Tailwind CSS, shadcn/ui
Developed an interactive platform with a vulnerability intelligence story, an AI Analyst interface, and a predictive Impact Dashboard.
Infrastructure — Vultr, Caddy, systemd, pnpm
Deployed the frontend and FastAPI backend on Vultr, with reverse proxy routing, automated release builds, health checks, and rollback support.
Challenges we ran into
- Extreme class imbalance: Only 85 of 64,120 training records were subsequently added to KEV. We used class weighting and ranking-based evaluation to optimize for identifying exploited vulnerabilities within a fixed remediation budget.
- Temporal data leakage: Vulnerability intelligence evolves over time. We engineered features using publication-time information and evaluated models chronologically to better represent real-world deployment.
- Integrating unstructured threat intelligence: We converted vulnerability descriptions into IBM Granite-derived text features and precomputed scores to reduce repeated inference costs.
- Reliable AI integration: We constrained the AI Analyst to predefined SQL queries and implemented rate limits and fallback logic to maintain predictable operation under watsonx.ai API limits.
Accomplishments that we're proud of
Improved identification of exploited vulnerabilities
In our chronological 2024-to-2025 backtest, PatchRadar identified 19 exploited CVEs versus 14 for the baseline within the same top-243 ranking cutoff, a 35.7% improvement in identified cases. This difference is promising but not statistically conclusive given the small positive sample.
In repeated grouped cross-validation, the model averaged 36.4 identified exploited CVEs versus 22 for the baseline within the top 442.
Built an end-to-end predictive intelligence system
Delivered a reproducible ML pipeline integrated with a production-deployed React application, FastAPI service, and IBM watsonx.ai-powered analytics interface.
Integrated IBM Granite into predictive modeling
Used Granite-derived semantic features alongside structured vulnerability indicators, demonstrating how foundation-model outputs can support downstream machine learning rather than serving only as a conversational interface.
Established a reproducible evaluation workflow
Implemented multiple historical validation approaches, model ensembling, precomputed features, and version-controlled prediction outputs to support repeatability and transparent performance comparisons.
What we learned
- Effective vulnerability management is a ranking problem. For organizations with limited remediation capacity, precision within the highest-priority vulnerabilities is more actionable than overall classification accuracy.
- Structured and unstructured signals are complementary. Combining engineered exploitation indicators with IBM Granite-derived text representations provides a richer basis for predictive analysis.
- Evaluation methodology is critical. Temporal leakage, class imbalance, and small positive samples can make model improvements appear more reliable than they are.
- Production AI requires controlled execution. Grounded SQL queries, deterministic fallbacks, and inference limits improve reliability and operational predictability.
What's next for PatchRadar
- Real-time intelligence ingestion: Automate CVE, EPSS, and KEV updates to support continuous vulnerability monitoring.
- Unified risk intelligence: Connect the AI Analyst directly to the predictive ranking engine for consistent, model-driven recommendations.
- Prediction explainability: Surface SHAP-based feature contributions for individual CVEs to support analyst review and remediation decisions.
- Continuous evaluation: Track future KEV additions to measure out-of-sample performance and recalibrate the ranking model.
- Enterprise integration: Connect PatchRadar to vulnerability management and ticketing platforms to convert risk predictions into actionable remediation workflows.
Our goal is to move vulnerability management from reactive severity-based triage toward proactive, evidence-driven risk prioritization.
Built With
- caddy
- css
- duckdb
- fastapi
- git
- granite
- ibm
- lightgbm
- python
- react
- scikit-learn
- shadcn/ui
- tailwind
- typescript
- vultr
- watsonx.ai