Visionflow default presentation

VisionFlow is Spatial Intelligence for the Age of Agents

• AI Symposium 2025 | Previously presented at AI Cafe v6 | IRIS ingenuity

---

The Problem: AI Is Making Knowledge Work Harder

---

A third of the global workforce are knowledge workers

• INGEST - REASON - WRITE - VERIFY — the universal loop
• $5 Trillion market addressable by agentic systems
• 72% of enterprises plan to deploy AI agents by 2026 — but fewer than 10% have succeeded (Gartner)

---

Current AI tools might have an interface problem

• Chat windows. Terminals. Inline completions. Flat text everywhere. Nobody seems to have insight yet as to how to fix this.
• 62% of workers report burnout from AI-augmented workflows

“AI introduced a new rhythm in which workers managed several active threads at once: manually writing code while AI generated an alternative version, running multiple agents in parallel, or reviving long-deferred tasks because AI could ‘handle them’ in the background. They did this, in part, because they felt they had a ‘partner’ that could help them move through their workload. While this sense of having a ‘partner’ enabled a feeling of momentum, the reality was a continual switching of attention, frequent checking of AI outputs, and a growing number of open tasks. This created cognitive load and a sense of always juggling, even as the work felt productive.”

• Harvard Business Review: AI Doesn’t Reduce Work, It Intensifies It

---

Agentic Engineering: How VisionFlow was Built

---

Claude Code is already 4% of all public GitHub commits — 135,000+ per day

• Projected to reach 20%+ of all daily commits by end of 2026
• Anthropic hit $1B annualised revenue in six months, faster than ChatGPT

---

VisionFlow is built with and for this paradigm

• Multi-agent Docker container with 5 AI personas (Claude, Gemini, OpenAI, Z.AI, DeepSeek) under supervisord
• 101 agent skills from architecture design to Blender scenes to hardware schematics
• Agentic development is not a feature, it’s the foundation
• 930+ source files, 18 months of continuous agentic engineering

---

What is VisionFlow / IRIS?

---

Powerful and secure collaborative agent orchestration platform

VisionFlow rendering a live ontology fashion & garment domain

• Thousands of nodes laid out by semantic physics — hierarchy, disjointness, and equivalence rendered as physical forces in real time

---

Ontology-Driven Semantic Physics

---

Why ontologies matter for AI agents

• Ontologies boost LLM performance: +55% fact recall, +40% correctness, 4.2x accuracy on schema queries (EMNLP 2025, data.world 2024)

---

Axiom-to-Force Translation — the core idea

• Formal ontology axioms interpreted as physical forces in a GPU simulation

OWL Axiom	Physical Force	Visual Result
SubClassOf(A, B)	Spring attraction (Hooke’s law)	Hierarchies form concentric shells
DisjointWith(A, B)	Coulomb repulsion	Contradictions create visible gaps
EquivalentClasses(A, B)	Strong spring (near-zero rest length)	Synonyms merge
Inferred axioms	Same laws at 0.3x strength	Prevents over-determination

• You can see the shape of knowledge — without reading a single label
• Academic citations
  • OG-RAG (EMNLP 2025): Ontology-grounded RAG improves fact recall by 55%, response correctness by 40%. aclanthology
  • GenAI Benchmark II (data.world, 2024): Knowledge graphs with ontology-based query checks yield 4.2x higher LLM response accuracy. data.world
  • OntoRAG (OpenReview): Enhances RAG by retrieving taxonomical knowledge from ontologies, suppressing hallucinations. openreview
• GPU implementation detail
  • 11 CUDA .cu files, 6,400+ lines, 28 __global__ kernels
  • Grid-accelerated N-body: spatial hash grid reduces O(N^2) to O(N * avg_neighbours)
  • Structure-of-Arrays memory layout for maximum GPU coalescing
  • Stability detection via parallel kinetic energy reduction — physics pauses when equilibrium reached
  • Whelk-rs: OWL 2 EL reasoner, <2s on 900+ classes, 90x LRU cache speedup

---

Semantic Physics to Visualise Meaning

---

Axiom-driven forces create spatial structure you can read

• Clusters form naturally from the ontology — no manual layout, no configuration — the shape is the meaning

---

Super Smooth Rendering - Precise Agent Control

---

Teams of humans, and teams of agents, in more natural spaces

• Transforms static documents into living knowledge ecosystems
• Ingests ontologies from Logseq notebooks via GitHub, reasons over them with OWL 2 EL inference, renders as interactive 3D graph with semantic physics
• Agents continuously analyse the graph, propose new knowledge via GitHub PRs, and respond to voice commands

---

Case Study: Creative Sprint at THG Ingenuity Studios

---

World Record for Longest AI Assisted Catwalk

---

From a standing start — 3 hours, on-site, zero preparation

• No pre-built workflows. No pre-existing assets, except this one image

  

• Provided screenshots of an existing workflow, then gave IRIS voice instructions. Everything autonomously produced.

• The human creative director never opened ComfyUI, Photoshop, or any creative application

• All workflows generated from scratch by AI agents

• Full report
  • https://github.com/DreamLab-AI/THG-world-record-attempt/blob/main/reports/world-record-report.pdf

• Full report
  • https://github.com/DreamLab-AI/THG-world-record-attempt/blob/main/reports/world-record-report.pdf

---

Case Study: The Results

Metric	Traditional Baseline	IRIS System	Multiplier
Per-image generation	15 min	~1 min	15x faster
Full campaign (36 assets)	~4 hours	30 min	8x faster
Extended campaign (138 assets)	~15+ hours	~96 min	~9.4x faster
Pipeline construction	4 hours (manual)	0 hours (autonomous)	Eliminated
Human input required	Continuous expert operation	21 voice prompts (~5 min)	1:19 ratio
Creative concepts	6-8 per session	45+ variations	5.6x more

• Pipeline construction: eliminated. The agents built the ComfyUI workflows, configured multi-GPU rendering, integrated APIs, ran QA — all autonomously.

---

21 voice prompts. 138 assets. That’s the ratio.

---

Media Co-Creation: What Agents Can Make

---

ComfyUI — autonomous creative pipelines

• Agents build, test, and run ComfyUI workflows without human intervention
• Multi-GPU pipeline configuration, API integration, quality assurance — all agentic

---

ANYTHING YOU CAN DO IN COMFYUI !!

---

Physically Based Textures from BIM (Revit)

---

Blender MCP — headless 3D creation via voice

First Blender test — headless container, returned the PNG

• “Connect to the Blender MCP and create me a swarm of shuriken with flocking behaviour. Use neural enhancements to test the swarming code using algorithmic breeding, then convert to Python for the remote MCP. Make 200 shuriken items black glass, each spinning on its central axis.”

A modern interpretation of Hypnerotomachia Poliphili (1499)

• Task(Initialize Hive Mind) ☐ Initialize Blender project with proper scene settings and units (feet) ☐ Create base terrain: Valley with sheer mountain cliffs using displacement ☐ Model Great Pyramidal Gate base plynth (1536ft x 1536ft x 35ft) ☐ Create pyramid body with 1410 parametric steps and internal staircase ☐ Configure dreamlike lighting with low perpetual sun and dramatic shadows ☐ Design kinetic bio-mechanical Medusa iris entrance system ☐ Apply white engineered surface material with fiber-optic seams to pyramid ☐ Create checkered marble courtyard floor (vast geometric grid) ☐ Model colossal winged horse in Corten steel/carbon fiber composite ☐ Create hollow elephant with terrazzo material and gold/silver flakes ☐ Build interactive 64-square chessboard with light panels (24ft x 24ft) ☐ Design elephant interior with sepulcher and royal statues ☐ Generate parametric golden lattice canopy structure ☐ Create kinetic Three Graces fountain with multi-tiered water system ☐ Arrange all assets in proper spatial relationships and optimize scene

---

More Things It Has Made

---

”Make me a pre-amp” — KiCad + ngspice via MCP

• “A preamp with a bit of character, not too expensive, nothing too flashy”
• 500-series “Character Toolbox” mic preamp — OPA1612, transformer saturation, JFET harmonics
• 32 components, $102 BOM, 95% production-ready — from a single voice prompt
• Pre-amp BOM detail
  • Component cost: $102.20|Manufacturing:$162.20 | Target price: $399-499 | Margin: 47.6-67.5%
  • 3x Cinemag CMMI-8-PCA transformers, OPA1612 dual op-amp, 2N5457 JFET, professional XLR connectors
  • MCP tools used: kicad.create_project, kicad.netlist_extraction, kicad.circuit_pattern_recognition, kicad.run_drc, kicad.generate_bom

---

Company website — auto-pushed to GitHub Pages

• DreamLab AI Consulting Ltd. — agents designed, built, and deployed it

---

CAVE System Quote — 300 pages in 4 hours

• Three-tier immersive system quote with HVAC specs, team selection, branding
• Selected team and branding guidelines autonomously from the DreamLab website
• CAVE quote detail

---

Industry reports — agents author formal documents autonomously

---

Capabilities

---

5-Tier Memory Architecture

Tier	Technology	Purpose
Ontology	OWL 2 EL + Whelk-rs	Semantic reasoning, inference
Knowledge Graph	Neo4j 5.13	Structured relationships, Cypher queries
Vector Memory	Qdrant	Semantic similarity search
Document RAG	GraphRAG / RAGFlow	Thousands of documents and books
Session Memory	Per-project agent context	Task continuity across sessions

• GitHub markdown (human-readable) as the single source of truth baseline

---

101 Agent Skills across 8 domains

• AI & Reasoning | Development & Quality | Agent Orchestration | Knowledge & Ontology | Creative & Media | Infrastructure | Document Processing | Architecture

---

Voice Interaction — 4-plane spatial audio model

• Private command (PTT) → Agent response (spatial) → Public voice (LiveKit SFU) → Public agent (spatial)

• <500ms command acknowledgement | Opus 48kHz | HRTF spatial panning | Unique voice per agent

• Agent skills list

  • AI & Reasoning: deepseek-reasoning perplexity perplexity-research pytorch-ml reasoningbank-intelligence
  • Development & Quality: build-with-quality rust-development pair-programming agentic-qe github-code-review
  • Agent Orchestration: hive-mind-advanced swarm-advanced swarm-orchestration flow-nexus-neural agentic-lightning
  • Knowledge & Ontology: ontology-core ontology-enrich import-to-ontology logseq-formatted docs-alignment
  • Creative & Media: blender comfyui comfyui-3d canvas-design ffmpeg-processing algorithmic-art
  • Infrastructure: docker-manager docker-orchestrator kubernetes-ops linux-admin infrastructure-manager
  • Document Processing: latex-documents docx xlsx pptx pdf text-processing
  • Architecture: sparc-methodology prd2build wardley-maps mcp-builder v3-ddd-architecture

• MCP ontology tools

  • 7 tools exposed via Model Context Protocol for AI agent read/write access to the knowledge graph:

  Tool	Purpose
  ontology_discover	Semantic keyword search with Whelk inference expansion
  ontology_read	Enriched note with axioms, relationships, schema context
  ontology_query	Validated Cypher execution with schema-aware label checking
  ontology_traverse	BFS graph traversal from starting IRI
  ontology_propose	Create/amend notes - consistency check - GitHub PR
  ontology_validate	Axiom consistency check against Whelk reasoner
  ontology_status	Service health and statistics

  • The Loop: Agent discovers - reads enriched context - proposes amendment - Whelk consistency check - GitHub PR - human review - merge - auto-sync - Whelk re-reasons - GPU re-layouts

• Voice architecture

  Plane	Direction	Scope	Trigger
  1	User mic - turbo-whisper STT - Agent	Private	PTT held
  2	Agent - Kokoro TTS - User ear	Private	Agent responds
  3	User mic - LiveKit SFU - All users	Public (spatial)	PTT released
  4	Agent TTS - LiveKit - All users	Public (spatial)	Agent configured public

  • Latency budget: STT 300ms + parse 1ms + agent 50ms + ACK 5ms = 410ms measured

---

Architecture

Layer	Technology	Detail
Backend	Rust 1.75+, Actix-web	373 files, 168K LOC, hexagonal architecture
Frontend	React 19, Three.js, R3F	377 files, 26K LOC, TypeScript 5.9
Graph DB	Neo4j 5.13	Cypher queries, bolt protocol
GPU	CUDA 12.4	28 kernels, 6,400+ lines across 11 files
Ontology	OWL 2 EL, Whelk-rs	EL++ subsumption, consistency checking
XR	WebXR, @react-three/xr	Meta Quest 3, hand tracking
Multi-User	Vircadia World Server	Avatar sync, spatial audio, entity CRUD
Voice	LiveKit SFU	turbo-whisper STT, Kokoro TTS, Opus codec
Protocol	Binary V3	21 bytes/node, delta encoding, 80% bandwidth reduction
Auth	Nostr NIP-07	Browser extension signing, relay integration
Agents	MCP, Claude-Flow	101 skills, 7 ontology tools, 5 AI backends

• Performance benchmarks

  Metric	Result
  Max nodes at 60 FPS	180,000 (RTX 3080)
  GPU vs CPU speedup	55x
  WebSocket bandwidth reduction	80% (Binary V3 vs JSON)
  Concurrent users	250+
  Ontology reasoning	<2s cold, 22ms cached
  Voice command ACK	410ms measured
  Cold start to interactive	~5.5s

• Development timeline

  Phase	Period	Milestone
  Prototype	Aug 2024	Working nodes and edges in Three.js
  Rust rewrite	Oct-Dec 2024	Actix actor system, Neo4j, hexagonal architecture
  GPU physics	Jan-Mar 2025	CUDA force-directed layout, 55x speedup
  Ontology engine	Apr-Jun 2025	Whelk-rs OWL 2 EL, axiom-to-force pipeline
  Multi-user + XR	Jul-Sep 2025	Vircadia, WebXR Quest 3, spatial voice
  Agent tools	Oct-Dec 2025	MCP ontology tools, GitHub PR loop, 101 skills
  IRIS studio	Jan-Feb 2026	Immersive configuration, settings hardening

---

Open Source and Why It Matters

---

This isn’t a product pitch — the software is free and open source

• DreamLab-AI/VisionFlow — MPL 2.0 licence

---

Why I built it

• My knowledge base moves with the evolving edge of the agentic technology stack
• It’s unbounded by frameworks which age quickly in this moment
• I get value from the development process, the tooling, and the knowledge management
• This informs my teaching, consulting, and training

---

Why it matters for you

• The 80/10 gap is real: 80% of enterprises want multi-agent orchestration, fewer than 10% have achieved it
• Current tools lack the structural foundation to coordinate agents meaningfully
• Ontology-grounded spatial workspaces provide that foundation
• Other people decide what to do with the cheat codes I have assembled

---

What you can do with it

• Media co-creation — Blender, ComfyUI, creative pipelines via voice
• Knowledge management — structured, reasoned, auditable
• Agent orchestration — 101 skills, 5 AI backends, spatial coordination
• Training platform — learn agentic development by using it

---

Where to find out more…

• Open source author: Disruption and Convergence Book (2022, arXiv)

  

  

Introduction to me

About Me

Chief Hallucination Officer DreamLab AI Consulting Ltd,

residential training in Eskdale.

Associate Director R&D dreamlab

.

Support learning for Saïd Business School

Full portfolio of online AI modules.

Founding member Agentic Alliance

.Link to original

---

Core Offering

<<<<<<< HEAD:workingGraph/pages/Visionflow default presentation.md

• DreamLab AI Consulting Ltd. Bespoke Product, agentic AI research, consulting and residential training.

=======

• DreamLab AI Consulting Ltd. Bespoke Product, agentic AI research, consultant & trainer

96bfccd33e14f89defe28b7ea96cad1e212cf040:workingGraph/pages/AI Symposium Presentation.md

---

Links and Resources

---

Source Code

• GitHub: DreamLab-AI/VisionFlow — Open source, MPL 2.0

---

Live Demos

• Ontology Explorer — Interactive 2D knowledge graph
• DreamLab AI — Company website (agent-built)

---

References

• The Original Book (2022) — Where it all started

• HBR: AI Doesn’t Reduce Work — The interface problem

• SemiAnalysis: Claude Code Inflection — Agentic development market

• OG-RAG Paper (EMNLP 2025) — Ontology performance evidence

• Videos

---

Live Ontology Explorer