Engineering

Things here don't fit on the homepage: architecture, correctness work, and numbers I actually measured. No adjectives, only mechanisms and tests.

Architecture

LEXORA

Document
  ↓
Ingestion
  ↓
Celery
  ↓
Chunking
  ↓
Embeddings
  ↓
PostgreSQL + per-user FAISS
  ↓
Retrieval + rerank
  ↓
LLM
  ↓
Citations

Lexora's pgvector migration is planned; per-user FAISS is the current, documented trade-off.

Measured numbers

• Lexora: 35 tests; retrieval recall probe 3/3 at top-3.
• vpnbox: v1.0.0-beta4, one stdlib-only Python file with no claimed numbers of its own. It measures live ping/speed of the relays before trusting them.

Full context lives in the project case studies.