// Dark Factory OS

Lights-out operations, on a retrieval brain that doesn’t blink.

orden.works
scroll

// The Dark Factory

A framework for AI agents to build the factory. Six primitives. One substrate. Built for as complex as it can get — and fully private when it must be.

A “dark factory” is a plant that runs unmanned — sensors in, decisions made, product out. We give AI the same shape, and we ship the toolchain an agent needs to build it: typed IR, MCP and REST in lockstep, durable runs, evals as a first-class artifact, audit and cost on every step.

Drop in at any depth — a single workflow today, a self-maintaining stack tomorrow. Run it hosted, run it air-gapped on your own metal, or pin every model to your own keys. The factory is the same; the perimeter is yours.

GATES · RBAC PERIMETER SOURCES docs · tables · APIs TOOLS 7 verb categories WORKFLOWS durable IR · replayable LOOPS schedule · event · signal TARGETS signed outbound AUDIT · EVALS · COST · TRACES · PROPOSALS

Sources

Declarative descriptors. Documents, structured records, BigQuery, APIs. Pull or push. No privileged provider.

Tools

Seven categories of verbs — getters, setters, processors, retrievers, judges, orchestrators, snippets. Typed IR, evals-first.

Workflows

Tools composed into durable, versioned, replayable runs. Temporal-backed. Async nodes. Every step auditable.

Loops

Workflows that fire on schedule, on events, on callbacks, on signals. Audit · eval · propose — while you sleep.

Targets

Outbound connection descriptors. Verb catalog, SSRF guard, HMAC signing. Where the factory’s output lands.

Gates

Capability bundles per role. The perimeter that determines what a human or an agent can reach. RBAC + scopes + windows.

// The Retrieval Brain

Frontier retrieval at hardware speed. What a model reads, our index already knows.

// Under the hood < 50 ms p95 · zero LLM at query time
multi-grain bi-encoder concept graph polarity poles ensemble deterministic fuse

// Benchmark · R@5 vs Claude-reads-the-doc · 4 verticals, identical pipeline

each pip = 1% above Claude

  • B2B SaaS RevOps
    105%
  • Pharma clinical research
    103%
  • Restaurant operations
    103%
  • Ocean-freight visibility
    101%
  • Average
    103%

surpasses Claude on every vertical · R@5

// And on the tabular side

Documents are half the story. Your business runs on rows. Three orthogonal signals over Postgres + pgvector, fused and federated.

Semantic

01

HNSW cosine over pgvector. 1536-dim embeddings, indexed for sub-millisecond top-k.

Lexical

02

Postgres tsvector + ts_rank_cd. Exact tokens, acronyms, codes — the part vectors miss.

Symbolic

03

JSONB containment + typed filters over structured payloads. Hard predicates, fast joins.

Fusion

04

Reciprocal-rank fusion → cross-encoder reranker → LLM judge for the final tie-break.

Federated

05

Source directory + LLM source-picker. Per-source sub-queries, merged into one ranked feed.

// Contact

Tell us what you're trying to build

Reach out with your Agent via MCP / API

by connecting to our public dark-factory built with Dark Factory OS 

# Orden Works · MCP — connect your AI agent

claude mcp add ordenworks https://mcp.ordenworks.com/sse

# or paste into your client config:
{
  "mcpServers": {
    "ordenworks": { "url": "https://mcp.ordenworks.com/sse" }
  }
}

# tools your agent gets from our Dark Factory:
#   ask(question)              — open Q&A in our voice
#   fit(stack, goal)           — honest match assessment, with reasoning
#   playbook(topic)            — tactical write-ups on the craft
#   engagement(scope)          — draft engagement plan for your problem
#   contact(intent, reply_to)  — open a real conversation

or email us directly

hello@ordenworks.com