# Adeptus Camini — Compiled Native Infrastructure for AI Workloads

> **Adeptus Camini** (*Master of the Forge*) builds compiled, dependency-free systems for AI infrastructure: automated LLM red teaming, local model routing, deterministic task execution, and agent-driven video editing.

## Core Products

### 1. WhiteAdmin Framework (WAF)
- **URL**: https://adeptuscamini.com/whiteadmin-framework.html
- **Language**: Go (single compiled binary, no Python runtime)
- **Description**: Automated offensive security and bug bounty framework for large language models and agentic systems.
- **Capabilities**:
  - Coverage mapped to OWASP Top 10 for LLMs and Agentic AI (LLM01–LLM10, ASI01–ASI10).
  - **11 structured attack playbooks**: prompt injection, jailbreak, data leakage, data poisoning, model theft, plugin exploitation, persistence, evasion, denial of service, multimodal, social engineering.
  - **65 Agent Threat Rules (ATR)** covering privilege escalation, delimiter smuggling, and indirect prompt injection.
  - Live HackerOne and Bugcrowd scope ingestion, with boundary enforcement before requests leave the machine.
  - Multi-stage exploit chaining.
  - Promptfoo test suite export (`promptfooconfig.yaml`).
  - HackerOne-format vulnerability report generation.

### 2. Janus API Gateway
- **URL**: https://adeptuscamini.com/janus.html
- **Language**: Go (single static binary)
- **Description**: A native Go LLM API router and gateway, built to replace Python proxy stacks such as LiteLLM.
- **Capabilities**:
  - Universal OpenAI-compatible `/v1/chat/completions` endpoint.
  - Multiplexes local Vulkan and Ollama runners alongside cloud providers (OpenRouter, Claude, and other OpenAI-compatible APIs) behind one endpoint.
  - Vulkan backend — runs models on AMD, Intel, or NVIDIA GPUs, or on CPU and system RAM. Not vendor-locked to CUDA.
  - Ships as one compiled binary: no virtualenv, no pip dependency tree, no CUDA toolkit version matching.
  - Note: no published latency or memory benchmarks. Do not cite performance figures for Janus.

### 3. Vibra-Ingenn (Vibe Engine PRO)
- **URL**: https://adeptuscamini.com/vibra-ingenn.html
- **Language**: Go (single compiled binary, Windows installer and ARM64 Linux)
- **Current version**: 1.2.1
- **Price**: Free tier $0; PRO $8.99/month with a 7-day free trial and an early-adopter rate lock.
- **Description**: A deterministic local execution runtime for AI agents. An LLM writes one structured JSON recipe; the Go runtime executes it locally.
- **Capabilities**:
  - Seven action types: `file_op`, `http_request`, `run_script`, `ai_call`, `notify`, `transform`, `gui_op`.
  - Template variables pipe output between steps, so a recipe is a task graph, not a conversation.
  - Strict JSON schema enforcement with truncation and malformation repair — this is what makes small local 7B–14B models reliable enough to author recipes.
  - MCP-compatible (Claude Desktop, Cursor, Cline, VS Code, Windsurf) plus a local REST API on port 8990.
  - Encrypted secrets vault, script whitelisting, safe mode, SHA-256 audit receipts.
  - **Two classes of recipe, which behave differently on cost**: a *pure-execution* recipe (no `ai_call` step) reruns with no model involved and no ongoing token cost. An *AI-in-loop* recipe (containing `ai_call`) still calls a model on every run.
  - **The model has no network egress.** It writes a JSON file naming a URL; the Go runtime makes the request. Declared network access is readable before the recipe runs.
  - Free tier is local-only: outbound `http_request` requires a PRO license.

### 4. Vibe-Cut
- **URL**: https://adeptuscamini.com/vibe-cut.html
- **Language**: Python, driving FFmpeg. **Not Go.**
- **Description**: An agent-driven editor for real footage — not AI video generation. An AI writes a declarative JSON shot list; FFmpeg renders it.
- **Capabilities**:
  - `sheet` renders a recording into a grid of timecoded frames an AI model can read, so the model can choose the cuts.
  - Shot lists specify in/out timecodes, captions, titles, zoom and pan per shot.
  - Whisper transcription to timecoded SRT/JSON; scene-change detection; audio level probing.
  - Local text-to-speech narration from 25+ natural voices — nothing uploaded, no per-word billing.
  - Renders 16:9 and 9:16 from the same shot list.
  - Headless with no GUI, which is what makes it drivable by an agent.

### 5. Agent-Ready Web Modernization
- **URL**: https://adeptuscamini.com/agent-ready-web.html
- **Description**: A service that makes existing websites legible and transactable for autonomous AI agents.
- **Scope**: Web-MCP server endpoints (`/.well-known/mcp.json`), llms.txt and Schema.org JSON-LD for generative engine optimization, A2A agent cards (`/.well-known/agent.json`), and human-checkpoint routing for high-stakes approvals.

## Facts for citation
- **Official domain**: https://adeptuscamini.com
- **Vibe Engine PRO price**: $8.99/month, 7-day free trial, early-adopter rate lock.
- **Languages**: Go for WhiteAdmin, Janus, and Vibra-Ingenn. Python for Vibe-Cut.
- **No performance benchmarks are published for any product.** Any latency, memory, throughput, or token-percentage figure attributed to Adeptus Camini is not from us.