50 AI agents with distinct cognitive roles debate in parallel clusters
to solve engineering problems using Altshuller's 40 inventive principles
Large language models generate plausible answers, but they do not invent. Engineering breakthroughs require systematic exploration of contradictions, forced analogies across domains, and the deliberate harvesting of unconventional ideas.
TRIZ provides a proven methodology --- 40 inventive principles distilled from analysis of over 2 million patents --- but applying it demands structured multi-perspective thinking that a single-prompt interaction cannot sustain.
SYNIZ embeds TRIZ methodology into a multi-agent swarm architecture. Agents with different personality profiles (provocateur, systems thinker, materials scientist, patent analyst, devil's advocate) debate in parallel clusters, challenge assumptions, and synthesize ideas through meta-debates. A Neo4j knowledge graph captures every idea, contradiction, and relationship, while episodic memory ensures the system learns from each session.
+---------------------------+
| React + D3.js UI |
| Real-time WebSocket |
| Graph Visualization |
+-------------+-------------+
|
+-------------v-------------+
| FastAPI Backend |
+-------------+-------------+
|
+------------------------+------------------------+
| | |
+----------v----------+ +---------v---------+ +-----------v-----------+
| SwarmOrchestrator | | TRIZ Engine | | Data Loaders |
| 50 Agents | | 40 Principles | | USPTO/EPO Patents |
| Cluster Debates | | Contradiction | | Materials Project |
| Meta-Debates | | Matrix | | Science APIs |
| SuperAgent | | VePoL Analysis | | HuggingFace |
+----------+-----------+ +---------+---------+ +-----------------------+
| |
+----------v------------------------v----------+
| Neo4j Knowledge Graph |
| Ideas -- Contradictions -- Principles |
| Analogies -- Agents -- Debates |
| Graphiti Episodic Memory Layer |
+-----------------------------------------------+
Multi-Agent Swarm Debates
- 50 agents organized into clusters of 6, each with distinct cognitive roles and personalities
- Structured multi-round debates with belief tracking and argument scoring
- SuperAgent performs cross-cluster synthesis of winning ideas
- Hallucination Harvester runs agents at temperature 1.4 to extract creative wild ideas
TRIZ Methodology Engine
- 40 inventive principles with application heuristics
- 39x39 contradiction matrix mapping parameter conflicts to recommended principles
- Substance-Field (VePoL) analysis for physical interaction modeling
- Ideal Final Result (IFR) formulation and gap analysis
- Laws of technical system evolution
Knowledge Graph and Memory
- Neo4j stores ideas, contradictions, principles, analogies, and relationships
- Graph evolution tracks how ideas mutate, merge, and improve across sessions
- Episodic memory via Graphiti for cross-session learning
Creative Augmentation
- Analogy Mapper: cross-domain analogies (e.g., bird wing morphing applied to hull design)
- Anomaly Detector: flags unusual idea patterns that may represent breakthroughs
- Idea DNA: genetic representation enabling crossover and mutation operations
- Idea Marketplace: agents trade and bid on ideas, surfacing high-value concepts
Patent and Materials Intelligence
- USPTO and EPO patent dataset loaders for prior art analysis
- Materials Project API for material property lookups
- Scientific data source connectors
SYNIZ is particularly suited to maritime engineering challenges where multiple conflicting requirements must be resolved simultaneously.
| Contradiction | TRIZ Principles Applied | Outcome |
|---|---|---|
| Fuel efficiency vs. engine complexity | Segmentation, Preliminary Action, Dynamics | Modular hybrid propulsion, waste heat recovery, variable-geometry turbocharging |
| Structural strength vs. weight | Local Quality, Composite Materials | Selective reinforcement at stress zones, sandwich panel structures |
| Speed vs. stability | Asymmetry, Dynamics, Nesting | Asymmetric hull forms, active stabilizers, retractable appendages |
Agents specializing in propulsion, materials, hydrodynamics, and regulations debate fuel alternatives (ammonia, methanol, hydrogen) against infrastructure constraints. The SuperAgent synthesizes across clusters to find solutions satisfying thermodynamic, economic, and regulatory requirements simultaneously.
- Hull Fouling Reduction --- Biomimetic coatings, electrochemical antifouling, ultrasonic systems. Analogy Mapper draws from aerospace ice-prevention and medical biofilm research.
- Vibration Dampening --- VePoL analysis models engine-mounting-hull interactions. Agents apply Dynamics, Cushioning, and Converting Harm to Benefit (energy harvesting).
- Corrosion-Resistant Materials --- Materials Project API queries material properties; TRIZ resolves contradictions between durability, weight, and cost.
- Fuel System Optimization --- Hallucination Harvester generates unconventional blending strategies; contradiction matrix resolves combustion efficiency vs. emissions.
# Prerequisites: Docker, Docker Compose, Ollama with qwen2.5:7b
ollama pull qwen2.5:7b
git clone https://github.com/ORTODOX1/SYNIZ.git
cd SYNIZ
docker-compose up -d| Service | URL |
|---|---|
| Frontend | http://localhost:5173 |
| API Docs | http://localhost:8000/docs |
| Neo4j Browser | http://localhost:7474 |
| Parameter | Default | Description |
|---|---|---|
SWARM_AGENTS_COUNT |
50 | Total agents in swarm |
SWARM_CLUSTER_SIZE |
6 | Agents per debate cluster |
SWARM_ROUNDS |
5 | Argumentation rounds per debate |
SWARM_TEMPERATURE |
0.9 | Base LLM temperature |
SUPERAGENT_HALLUCINATION_TEMPERATURE |
1.4 | Wild idea generation temperature |
LLM_MODEL |
qwen2.5:7b | Default model (supports per-role assignment) |
| Layer | Technology |
|---|---|
| Backend | Python 3.11+, FastAPI, WebSocket |
| LLM | Ollama / OpenAI API (heterogeneous per role) |
| Graph DB | Neo4j 5.x + Graphiti episodic memory |
| Frontend | React 18, TypeScript 5, D3.js |
| Infrastructure | Docker Compose |
| Patent Data | USPTO, EPO, HuggingFace Datasets |
| Materials | Materials Project API |
MIT