← Back to Projects

DefenseAI

Live Demo

Key Design Principles

  • Human-in-the-Loop — Every AI output requires explicit human review and approval before becoming actionable
  • Safety-First Orchestration — Multi-stage validation pipeline in Rust rejects malformed, dangerous, or low-confidence outputs
  • Immutable Auditability — Every prompt, response, and decision is logged with SHA-256 cryptographic hashing for tamper detection
  • Separation of Concerns — Rust handles AI orchestration and safety; Django handles application logic, UI, and user management
  • Defense-Grade Reliability — Retry logic, timeout enforcement, structured error handling, and graceful degradation

Request Flow

  1. Analyst submits raw intelligence data through the Django web interface
  2. Django sends the data to the Rust Orchestrator via HTTP POST
  3. Rust constructs a structured prompt and calls the LLM API with retry/timeout logic
  4. The LLM response passes through the Safety Validation Pipeline (schema check, content safety, logical consistency, confidence scoring)
  5. Every step is logged in the Immutable Audit Trail with SHA-256 hashes
  6. The validated brief is returned to Django and placed in the Review Queue
  7. A Reviewer or Admin must explicitly Approve, Reject, or Request Revision

Rust Backend — AI Orchestrator

The Rust service is responsible for all AI-related operations. Rust was chosen for its memory safety, zero-cost abstractions, and strict type system — critical properties for a safety-critical system where the LLM is treated as an unreliable oracle.

Django Application — Web UI & Workflows

Django handles the user-facing application: authentication, role-based access control, brief lifecycle management, and the enterprise web interface. It communicates with the Rust Orchestrator via REST API and persists all application data in PostgreSQL.


Primary Model

  • Model: Qwen/Qwen3-32B-AWQ
  • Quantization: AWQ (4-bit)
  • Parameters: 32 billion
  • Serving: vLLM (OpenAI-compatible API)
  • Hardware: 2x NVIDIA RTX 3090 (48GB VRAM)
  • Timeout: 90 seconds per request
  • Max retries: 3
  • Temperature: 0.3 (low for structured output)

analyst:analyst123 - source data, view own briefs, view brief details

reviewer:reviewer123 - All Analyst permissions + approve, reject, or request revision on briefs


admin:admin123 - All Reviewer permissions + observability dashboard + Django admin panel

Gallery