Repository navigation
Home
An alert triage pipeline that measures how often its AI is wrong.
TriageLens assists security analysts with alert investigation, threat intelligence, event correlation, detection, and AI-assisted triage. Its goal is to measure model errors against independent ground truth while keeping people responsible for investigation and response.
Human oversight: AI assessments are investigative aids. Analysts retain decision-making authority, and ground truth remains independent of model predictions.
| Your goal | Guide |
|---|---|
| Install and run a synthetic example | Getting Started |
| Understand components and boundaries | Architecture |
| Understand model outputs and metrics | AI Triage and Evaluation |
| Investigate alerts and detections | Threat Intelligence and Detection |
| Run tests and contribute | Development and Testing |
flowchart LR
A[JSON / CSV / EVTX] --> B[Normalize and correlate]
B --> C[Threat intelligence]
C --> D[Sanitized Groq triage]
D --> E[Assessment and uncertainty]
L[Independent labels] --> F[Evaluation]
E --> F
B --> G[Sigma and hunting]
S[Registered file snapshots] --> Y[Optional YARA]
E --> H[Storage and API]
F --> H
G --> H
Y --> H
H --> I[Dashboard and analyst review]
The architecture guide explains component boundaries. Rule detections, AI predictions, independent labels, and analyst dispositions remain separate.
| Area | What TriageLens provides |
|---|---|
| Intake and timelines | JSON, CSV and EVTX intake, normalized alerts, provenance and time-windowed correlation. |
| Threat intelligence | VirusTotal, AbuseIPDB and URLhaus lookups, with explicit unavailable results. |
| AI triage | Groq assessments with disposition, severity, evidence references, uncertainty and investigation recommendations. |
| Detection | Curated Sigma subset, KQL-style hunting filters, optional YARA-X and ATT&CK annotations. |
| Evaluation | Confusion metrics, severity agreement, calibration, intervals and scoped coverage against independent labels. |
| Analyst workspace | FastAPI and React views for investigations, detections and separate human reviews. |
| Storage | SQLite by default; optional PostgreSQL, Docker Compose and a saved-hunt worker. |
- Alert Intake and Correlation — inputs, EVTX, timelines and provenance.
- Threat Intelligence and Detection — enrichment, rules, YARA and hunting.
- AI Triage and Evaluation — evidence, model outputs, scoring and formulas.
- API and Dashboard — endpoints and analyst workflows.
- Deployment and Operations — PostgreSQL, Compose, the worker and recovery.
- Development and Testing — offline tests and CI.
Groq is the supported runtime AI provider and SQLite is the default database. EVTX intake and detection features are implemented, but reviewed real-data evaluation and live quality acceptance of the current triage-v2 policy remain pending. Small synthetic samples do not establish real-world accuracy.
Intelligence failures do not establish that an indicator is safe, and structured AI output does not guarantee factual or operational correctness. Masking targets supported sensitive-data patterns but is not exhaustive; review telemetry before sending it to external services.
For status and evidence, see the task list and README.
Repository · Getting Started · Architecture decisions · MIT License