README.md

metadata

language:
  - en
license: apache-2.0
tags:
  - agent
  - multi-agent
  - langgraph
  - gguf
  - lora
  - qwen
  - prediction-markets
  - mad-gambit
  - solidity
  - tandem-core
  - agent-q3
  - agent-q30-smart
pipeline_tag: text-generation

Agent Q3

MAD Gambit — Multi-agent intelligence layer Contains both Agent Q3 (self-improving production orchestration and self-improving research pipelines)

---

Variant

Variant	Role	Key Tech
Agent Q3	Production orchestration	Tandem Core · ComputeRouter · LangGraph · MCP · Solidity audit
Self-improving research	LoRA fine-tuning · arXiv ingestion · ChromaDB · Unsloth · benchmarks
Agent Q3 Smart

---

File Structure

Agent-Q3/
├── README.md
│
├── hq/                          ← Agent Q3 — Production
│   ├── Dockerfile
│   ├── docker-compose.yml
│   ├── requirements.txt
│   ├── .env.example
│   ├── orchestrator.py          ← FastAPI + ComputeRouter
│   ├── compute_router.py        ← Multi-backend weighted routing
│   ├── tandem_core.py           ← Gemma4 Reasoner + Qwen3.5 Coder chain
│   ├── langgraph_graph.py       ← StateGraph agent orchestration
│   ├── mcp_config.json          ← 14 MCP server connectors
│   ├── litellm_config.yaml      ← LiteLLM routing config
│   └── metrics.py               ← Prometheus /metrics endpoint
│
├── Agent Q3 /                         ← Agent Q3 — Research
│   ├── Dockerfile
│   ├── requirements.txt
│   ├── .env.example
│   ├── training_pipeline.py     ← Unsloth LoRA fine-tuning
│   ├── arxiv_ingestor.py        ← arXiv fetch → chunk → embed
│   ├── chromadb_store.py        ← Vector store (nomic-embed-text 384-dim)
│   ├── feedback_collector.py    ← DPO/RLHF signal capture
│   ├── benchmark_runner.py      ← Domain QA evaluation
│   ├── lora_pusher.py           ← Push adapters to HF
│   └── langgraph_graph.py       ← StateGraph agent orchestration
│
└── shared/                      ← Shared utilities
    ├── auth.py
    ├── logger.py
    └── config.py

---

HQ — Quick Start

cd hq
cp .env.example .env
docker compose up --build

Endpoints:

• POST /v1/chat — auto-classify → Reasoner or Coder
• POST /v1/reason — force (planning, research, audit)
• POST /v1/code — force (code, debug, file ops)
• POST /v1/tandem — Gemma4 reasons → Qwen3.5 implements
• GET /health — backend health + loaded models
• GET /metrics — Prometheus metrics

Agent Q3 — Quick Start

cd agent=q3
cp .env.example .env
pip install -r requirements.txt
python training_pipeline.py

---

Linked HF Assets

Type	Label	Link
Model	Agent Q3 (unified)	madDegen/agent-q3-core
Dataset	Agent Q3 (unified)	madDegen/agent-q3
Space	Agent Q3 (unified)	madDegen/agent-q3-space

---

License

Apache 2.0