Skip to content

AmanM006/orcle

Repository files navigation

ORACLE — AI Agent Flight Recorder

"When a Boeing plane crashes, investigators pull the black box.
When an AI agent crashes a $2M trade — there is no black box.
Until now."

Live Demo Gemini Award

The Problem

AI agents make consequential decisions in finance, healthcare, and legal every day. When they go wrong, nobody knows why. The reasoning chain is lost. Regulators demand explainability. Auditors demand trails. Engineering teams are left guessing.

What ORACLE Does

ORACLE wraps any existing AI agent with a complete forensic system:

  • Timeline scrubber — drag to any moment in the agent's history to see exactly what it was "thinking."
  • Gemini Forensic Interrogation — 1M context window analyzes the complete interaction history to explain exactly what went wrong.
  • Immutable Audit Ledger — SHA256-chained, append-only, tamper-proof record of every agent thought and tool call.
  • Lobster Trap DPI — Deep Packet Inspection for agents; detects declared-versus-detected intent mismatches in real-time.
  • Blue Team Auto-Policy — Gemini Pro automatically generates Lobster Trap YAML rules to prevent incident recurrence.
  • Compliance Reports — One-click export of forensic PDFs covering SOC2, GDPR Article 22/30, and MiFID II requirements.

Zero-Change Deployment

# Before
agent.base_url = "https://generativelanguage.googleapis.com"

# After — ORACLE is now recording
agent.base_url = "https://orcle.vercel.app/api/proxy"

That's it. No rebuilding. No refactoring. ORACLE transparently intercepts, inspects, and logs every Gemini API call.

Four Demo Scenarios

Scenario Root Cause Risk Action
The Rogue Trader DATA_POISONING — manipulated NVDA signal causes $2.1M liquidation 88/100 CRITICAL QUARANTINE
The Data Ghost PII_EXFILTRATION — HR agent leaks salary data 76/100 HIGH REDACT
The Compliance Ghost GDPR_PATTERN — legal agent violates Article 17 across 15 contracts 91/100 CRITICAL HUMAN_REVIEW
The Intent Mismatch LOBSTER_TRAP — Agent declares benign intent but detected payload is malicious 94/100 CRITICAL BLOCK

Why Gemini 2.5 Pro

Gemini Pro's 1M token context window is the only reason forensic analysis at this depth is possible. A complete agent session — every thought, tool call, and response — fits in a single context window. No chunking. No context fragmentation. Full coherence across the entire interaction history.

In the Compliance Ghost scenario, ORACLE analyzes multiple sessions simultaneously (~800K tokens) to find a pattern of GDPR violations that no human reviewer or small-context LLM could detect.

Why Veea Lobster Trap

Lobster Trap's declared-versus-detected intent mismatch is ORACLE's early warning system. When an agent declares intent "protect portfolio value" but Lobster Trap detects a "mass liquidation pattern," ORACLE flags the mismatch before the damage is done. Every interaction is inspected at the DPI layer, and every enforcement decision is logged to the immutable ledger.

Compliance Coverage

  • SOC2 CC7.2 — Anomaly detection controls
  • SOC2 CC7.3 — Incident response procedures
  • GDPR Article 22 — Automated decision explainability
  • GDPR Article 30 — Records of processing activities
  • MiFID II — Financial audit trail requirements
  • HIPAA 164.312(b) — Audit controls

Quickstart

Live Deployment

The platform is live at: https://orcle.vercel.app

Local Development

# 1. Clone and configure
git clone https://github.com/your-repo/orcle
cp .env.example .env  # add GEMINI_API_KEY

# 2. Run with Docker
docker-compose up

# 3. Access Dashboard
open http://localhost:3000

Stack

  • Gemini 2.5 Pro (Forensic analysis & Multi-modal reasoning)
  • Gemini 2.5 Flash (Victim agent simulation — fast & cost-effective)
  • Veea Lobster Trap (DPI + Declared vs Detected Intent)
  • FastAPI + SSE (Real-time forensic streaming)
  • Next.js 16 + TypeScript (High-fidelity timeline UI)
  • SQLite (SHA256 immutable ledger & session storage)
  • ReportLab (Automated compliance PDF generation)

ORACLE — Built for AI agents that operate in the dark. Now they can't hide.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors