The Third Eye: VIntel Platform
A multi-agent cybersecurity intelligence platform that aggregates, curates, and ranks threat-intelligence news using LLM-powered editorial agents. The system employs a two-round deliberation model with Borda-count cross-story ranking to overcome grade inflation, backed by a polyglot microservice architecture spanning Go, TypeScript, Rust, and React. Built as a course project for IE 624 — Generative & Agentic AI under Prof. Manjesh K. Hanawal.
Overview
The cybersecurity landscape generates a flood of news, vulnerability disclosures, and threat reports daily. VIntel Platform tackles this information overload by deploying multiple LLM-powered agents that independently evaluate incoming stories and then engage in a structured deliberation process to produce reliable, consensus-based rankings.
The key innovation is a two-round editorial deliberation model. In the first round, each agent independently scores stories on relevance, severity, and novelty. In the second round, agents review each other's assessments and refine their scores. The final ranking uses Borda-count aggregation across agents, which inherently resists the grade inflation that plagues simple averaging of LLM-generated scores.
The platform includes a persistent memory layer so agents can recall prior evaluations and maintain context over time, and implements the open A2A (Agent-to-Agent) protocol for interoperability with external agent systems.
Architecture
VIntel follows a polyglot microservice architecture where each service is implemented in the language best suited for its role. The services communicate asynchronously and are designed for independent deployment and scaling.
- Ingestion Service (Go) — high-throughput collector that aggregates cybersecurity news from RSS feeds, APIs, and web scrapers. Go's concurrency model handles parallel fetches efficiently across dozens of sources.
- Agent Orchestrator (TypeScript) — manages the multi-agent editorial workflow, coordinating the two-round deliberation, dispatching stories, collecting scores, and running the Borda-count ranking algorithm.
- Memory Service (Rust) — provides a persistent memory layer for agents, storing prior evaluations, user preferences, and contextual data with low-latency access. Rust ensures memory safety and performance for this latency-critical component.
- Frontend (React) — presents the curated intelligence feed with ranked stories, agent consensus scores, and drill-down views into individual agent assessments.
- A2A Gateway — implements the open Agent-to-Agent protocol, allowing external agent systems to submit stories for evaluation or subscribe to the curated feed.
Key Features
Two-Round Deliberation
Multiple LLM agents independently evaluate each story in the first round, then review and refine their assessments in a second round. This mimics an editorial board's review process, producing more calibrated and reliable scores than single-pass evaluation.
Borda-Count Ranking
Instead of averaging raw scores (which suffer from LLM grade inflation), the system converts each agent's scores into ordinal rankings and applies Borda-count aggregation — producing a final ranking that reflects relative importance rather than inflated absolute scores.
Persistent Memory Layer
Agents maintain state across sessions through a dedicated memory service, enabling temporal awareness — recognizing follow-ups to prior incidents, adjusting scores to evolving threats, and avoiding redundant coverage of the same event from multiple sources.
A2A Protocol
The platform implements the open Agent-to-Agent protocol, enabling interoperability with external agent systems. Third-party agents can submit stories for evaluation, query the curated feed, or integrate VIntel's ranking into their own workflows through a standardized interface.