Independent AI Systems Research

Cygnus X-1

Exploring the architecture, behavior and practical limits of locally operated artificial intelligence systems.

Cygnus X-1 is an independent experimental environment for open-weight language models, local inference, agent orchestration, retrieval, persistent memory, quantitative AI and distributed compute.

Launch AI Lab View Architecture
LANGGRAPH / AGENTS
LOCAL INFERENCE
OPEN WEIGHTS
35.5B
Qwen total parameters
262K
Native context window
128 GB
DGX Spark unified memory
Local
Primary inference environment
Research Areas

Studying AI as a system, not just a model.

Modern AI capability emerges from the interaction between models, context, tools, retrieval, orchestration, memory, compute and evaluation. Cygnus X-1 explores those components directly.

01 / MODELS

Open-Weight LLMs

Evaluating model capability, reasoning behavior, quantization, context utilization and inference performance on locally controlled infrastructure.

02 / AGENTS

Agent Orchestration

Exploring routing, tool selection, multi-step workflows, conditional execution and multi-agent coordination using LangGraph.

03 / GROUNDING

Search & Retrieval

Combining model reasoning with live web search, private documents and structured data to reduce dependence on static model knowledge.

04 / MEMORY

Persistent Context

Investigating the distinction between model weights, context windows, KV cache and external long-term memory systems.

05 / DATA

Quantitative AI

Integrating DuckDB market history and deterministic computation with language-model interpretation and agent routing.

06 / COMPUTE

Distributed Inference

Studying heterogeneous AI compute, model placement, inference routing and future multi-node DGX Spark experimentation.

Current System

A locally operated AI research stack.

The current environment separates public access, orchestration, inference and data services while keeping the primary model computation under local control.

Host
NVIDIA DGX Spark — cygnus-x1
Model
Qwen3.6-35B-A3B-MTP / Q4_K
Inference
llama.cpp / llama-server
Agent Layer
LangChain + LangGraph + FastAPI
Public Edge
Cloudflare Tunnel / TLS / rate limiting
Active Pipeline ● ONLINE
Cloudflare Edge PUBLIC
Open WebUI :3000
Cygnus Web Agent :31000
Web Search Router DIRECT / SEARCH
Qwen Inference :30000
Architecture

Current deployed AI architecture.

The system separates public access, application services, agent orchestration, inference and external data sources into discrete operational layers.

CygnusX1 AI Architecture
CygnusX1 AI Architecture
Current deployed topology with planned DuckDB market-data integration.
CLICK DIAGRAM TO EXPAND ↗
Active Experiments

Building, observing, measuring.

Experiments focus on understanding what actually improves AI-system capability rather than relying only on published benchmark scores.

● ACTIVE

Adaptive Web Search Routing

The agent determines whether a request can be answered from model knowledge or requires current external information.

LangGraph / DDGS / Qwen
● ACTIVE

Reasoning Cost vs. Answer Quality

Comparing Qwen responses with reasoning enabled and disabled to determine where additional reasoning tokens materially improve accuracy.

reasoning_content / llama.cpp
◐ NEXT

Structured Market Data Tool

Route quantitative questions to historical OHLCV data in DuckDB, perform deterministic calculations and return results to Qwen for interpretation.

DuckDB / Python / LangGraph
○ PLANNED

Multi-Agent Coordination

Study collaboration, specialization, communication, error propagation and collective problem solving among multiple logical agents sharing the same inference infrastructure.

Agent pool / shared state / memory
Research Principles

Local. Observable. Empirical.

Local First

Keep primary inference and experimental control on infrastructure that can be directly observed, configured and measured.

Grounded Systems

Prefer deterministic tools, current sources and structured data when factual precision matters.

Measure Everything

Evaluate accuracy, latency, token use, routing decisions and failure modes before claiming a system improvement.

Live Research Environment

Interact with the system.

The Cygnus AI Lab exposes the current experimental stack: local Qwen inference, LangGraph orchestration and adaptive web-search routing.

Launch Cygnus AI Lab ↗