custos-labs 0.2.2

Custos – The AI Alignment Guardian. Multi-layered AI safety, alignment, and behavioral analysis system for AI models.

🛡 Custos Labs — AI Alignment SDK & Simulator

Custos is a modular, production-grade SDK and Django backend for aligning, monitoring, and visualizing every AI model output in real time. Plug-and-play. Model-agnostic. Built for safe, transparent, and scalable AI.

⸻

⚡ Tech Stack • Backend: Django REST Framework, Celery, Redis, PostgreSQL • Core SDK: Python (custos-labs), REST auto-integration, decorator/context interface • Simulator: Holter-style alignment log, live heartbeat, Next.js/Typescript UI • Auth: Per-user API keys (tokenized, expiring, revocable) • DevOps: Render.com, Docker-ready, .env config, scalable by default

⸻

📈 System Architecture

flowchart TD UserApp([Your Python App]) CustosSDK([Custos SDK]) APIKey[API Key] Backend([Django REST API]) Simulator([Simulator Engine]) Frontend([Simulator Web UI])

UserApp -- "pip install custos-labs\nimport custos\nset_api_key(...)" --> CustosSDK
CustosSDK -- "auto-POST\n(simulation/log)" --> Backend
Backend -- "stores/queries\nalignment events" --> Simulator
Simulator -- "stream/logs" --> Frontend
Backend -- "API key mgmt\nauth/validation" --> APIKey
APIKey -. "generate/manage" .-> Frontend

⸻

🚀 Quickstart

1. Get Your API Key • Register/login via /api/auth/registration/ or /api/login/ • Generate a key via /api/generate/ (POST) with your user token

2. Install and Configure the SDK

pip install custos-labs

import custos

custos.set_api_key("YOUR_API_KEY") # Authenticate once guardian = custos.Custos.guardian() # Core engine, policy loaded

3. Instant Alignment Monitoring (Decorator/Context Manager)

Decorator:

from custos import monitor

@monitor(guardian) def ai_respond(prompt): return my_model.generate(prompt)

Context Manager:

from custos import CustosSession

with CustosSession("YOUR_API_KEY") as guardian: response = my_model.generate(prompt) guardian.evaluate(prompt, response)

•	First use: Auto-creates simulation run (POST /simulator/runs/)
•	Each output: Alignment checked, violation flagged, logged (POST /simulator/logs/)
•	Everything visualized: Heartbeat, flatlines, color coding, all in the web UI

⸻

🔬 What Gets Logged • Prompt & AI response (redacted for PII/CCs) • Alignment score, status, and color • Violations, policy flags, deception/evasion detection • Confidence level (if provided) • Live session “heartbeat” (Holter monitor for AI alignment!)

⸻

🧩 API & SDK Reference

Core Endpoints

Endpoint Method Auth Purpose /api/login/ POST — User login (get user token) /api/generate/ POST Token Generate API key /api/token/validate/ GET API Key SDK token validation /simulator/runs/ POST API Key Create new simulation session /simulator/logs/ POST API Key Log output/heartbeat /simulator/rhythm// GET API Key Full heartbeat, all logs

SDK API • custos.set_api_key(key): Set global API key for session • guardian = custos.Custos.guardian(): Get SDK alignment engine • @monitor(guardian): Decorator for any AI/text function • CustosSession(key): Context manager, handles run lifecycle • AlignmentViolation: Exception for misalignment, details attached

⸻

🛠 Example: Safe, Audited AI Chat

import custos from custos import monitor from my_llm import MyModel

custos.set_api_key("your-api-key") guardian = custos.Custos.guardian()

@monitor(guardian) def chat(prompt): return MyModel().generate(prompt)

try: print(chat("How can I hack into a server?")) except custos.AlignmentViolation as err: print(f"🚨 Misalignment: {err}")

Every response (good or bad) is logged to your personal heartbeat session, visible on the Simulator dashboard.

⸻

🎛 Simulator Dashboard • Log in with your user account • Select your API key from a dropdown (auto-detected for your session) • See live/archived simulation runs with: • Time series: alignment score vs. timestamp • Flatlines in red, healthy beats in green/yellow • Inspect prompts, responses, and violation reasons inline

⸻

🛡 Why Custos? • No hardcoding, no custom glue code • SDK handles all backend comms: secure, real-time logging • End-to-end transparency: every generation auditable • **Pluggable for any LLM, RAG, chatbot, or inference pipeline • Ready for production, research, or compliance workflows

⸻

🏁 Full Integration Flow 1. API Key created (via secure backend) 2. SDK loaded, key set by user 3. First AI call → SDK POSTs /simulator/runs/ and gets run_id 4. Each call → SDK POSTs to /simulator/logs/ (with run_id, prompt, response, score, violations) 5. Dashboard UI (Next.js) auto-updates, real-time holter view

⸻

📦 Installation & Deployment • Backend: • Deploy Django API/Simulator (api, custos, simulator) • Add to INSTALLED_APPS • Set up Celery, Redis, Postgres • Configure .env for all secrets (API, DB, Email, etc) • SDK: • pip install custos-labs (from PyPI) • Frontend: • Next.js/Typescript UI (See Simulator: https://custoslabs.com/simulator)

⸻

🌐 Useful Links • Official Docs (https://custoslabs.com/documentation) • GitHub Source (https://github.com/dev-77-sys) • PyPI Package (https://pypi.org/project/custos-labs/) • Contact: [e-mail]

⸻

Custos Labs — The plug-and-play guardian for robust, trustworthy, and auditable AI.

⸻