Material KAI Vision Platform — Full Business & Product Overview

Confidential | For Investor & Partner Use Version 3.5 — March 2026

Executive Summary

Material KAI Vision Platform is an AI-powered B2B SaaS platform for the architecture, interior design, and construction materials industry. It transforms the way professionals discover, specify, visualize, and procure building materials — replacing static PDF catalogs and fragmented manual searches with a unified intelligent platform powered by 20+ AI models across 8 providers.

The platform serves 5,000+ active users across three professional groups: buyers (architects, designers, sourcing agents), suppliers (manufacturers, brands), and platform operations. It is live in production at materialshub.gr with 99.5%+ uptime, 10,000+ cataloged products, and 1,000+ processed PDFs.

Core value propositions:

For buyers: Find the exact material — by look, feel, specification, image, or natural language — in seconds instead of hours. Then visualize it in a 3D interior design, walk through a VR version of the room, and request a quote — all within one platform.
For suppliers/manufacturers: Automatically digitize your entire catalog from a single PDF upload. Get real-time visibility into who is searching for your products, what the market trends are, and receive qualified enquiries directly.
For the platform: A credit-based consumption model tied directly to AI feature usage, Stripe-powered with volume discounts, layered over a marketplace that creates strong network effects as both sides grow.

The Problem

The architecture, interior design, and construction materials industry is a massive, globally fragmented market still running largely on PDFs, trade fairs, and manual outreach.

Pain for buyers (architects, interior designers, sourcing agents):

Finding the right material means searching across hundreds of manufacturer PDFs, websites, and physical showrooms — each with different formats, terminology, and cataloging standards
No cross-manufacturer visual search exists: "find me something that looks like this image" is impossible in any current tool
Specification-based search ("slip-resistant matte ceramic, 30×60cm, suitable for external use") requires manual page-by-page reading
No structured workflow from inspiration → specification → quote → project tracking exists in one tool
Design visualization requires separate tools (Photoshop, Revit, SketchUp) — completely disconnected from the material catalog

Pain for manufacturers and suppliers:

PDF catalogs are invisible to search engines and AI systems
No visibility into buyer interest: who viewed what, which products are trending
No competitive price intelligence without manual monitoring
No structured digital channel for receiving and managing qualified enquiries
Maintaining a web presence requires constant manual effort per product

The Solution

Material KAI Vision Platform closes this gap with a fully integrated AI platform that handles the complete lifecycle:

Manufacturer uploads PDF → AI extracts, understands & indexes all products
          ↓
Buyer searches by text, image, color, texture, or specification
          ↓
AI agents explain, compare & recommend materials from the catalog
          ↓
Interior Designer agent generates room visualizations using those materials
          ↓
VR generation turns any design into a walkable 3D world (in-browser)
          ↓
Buyer saves materials to Moodboards → builds a Quote → tracks the project
          ↓
Manufacturer sees who is interested, manages enquiries, tracks market trends

Every step — ingestion, search, design, visualization, procurement — runs on the same platform, creating a flywheel: more suppliers → richer catalog → better search results → more buyers → more enquiries → more suppliers.

Platform Architecture

Layer	Technology	Hosting
Frontend	React 18 + TypeScript + Vite + Shadcn/UI	Vercel Edge Network (global CDN)
Backend API	FastAPI + Python 3.11 (MIVAA service)	Dedicated server (DigitalOcean)
Database	Supabase PostgreSQL 15 + pgvector 0.8.0	Supabase managed cloud
Edge Functions	30+ Deno/TypeScript functions	Supabase managed cloud

Production URLs:

Platform: materialshub.gr
API: v1api.materialshub.gr
API Docs: v1api.materialshub.gr/docs

Scale characteristics:

Frontend served via Vercel global edge — zero latency regardless of user geography
Backend handles unlimited async job queue with checkpoint-based recovery
Database uses connection pooling and optimized halfvec (float16) vector indexes
170+ API endpoints, 20 categories, 60+ Supabase Edge Functions, fully documented via OpenAPI/Swagger

AI & Intelligence Layer

The platform's core differentiator is not a single AI model — it is the orchestration of 20+ AI models deeply embedded into every layer: ingestion, enrichment, search, agent interactions, design generation, and analytics.

Full AI Model Stack

Provider	Model	Role	Cost
Anthropic	Claude Sonnet 4.5	Jarvis agent, deep product analysis, metadata extraction, B2B research	$3 input / $15 output per 1M tokens
Anthropic	Claude Haiku 4.5	Fast classification, content detection, B2B web search	$0.80 input / $4 output per 1M tokens
OpenAI	GPT-4o	Alternative product discovery, multimodal tasks	$2.50 input / $10 output per 1M tokens
OpenAI	GPT-4o-mini	Query intent parsing, lightweight operations	$0.15 input / $0.60 output per 1M tokens
Voyage AI	voyage-3.5 (1024D)	Primary text embeddings + understanding embeddings	$0.06 per 1M tokens
HuggingFace	Qwen3-VL 32B Vision	Image analysis, OCR, material recognition (69.4% MMMU, #1 OCR benchmark)	Cloud endpoint
HuggingFace	SigLIP2 (768D × 5 types)	Visual / color / texture / style / material embeddings	Cloud endpoint
Replicate	FLUX.1-dev, FLUX.1-schnell, SDXL, SD3, Playground v2.5, Kandinsky 2.2, Proteus v0.2	Text-to-image interior design generation	Per image
Replicate	ComfyUI Interior Remodel, Interiorly Gen1 Dev, Designer Architecture + 4 others	Image-to-image interior transformation	Per image
Replicate	proplabs/virtual-staging	AI room staging from empty photos	20 credits/run
Replicate	Wan 2.1 i2v 720p, Runway Gen4 Turbo	Interior video generation (budget + premium)	12/40 credits
Replicate	meta/sam-2, ali-vilab/anydoor	Pixel-precise mask generation, product placement	Per use
WorldLabs	Marble mini + plus	3D Gaussian Splat VR world generation	50/200 credits per world
Google Gemini	gemini-3.1-flash-image-preview, gemini-3-pro-image-preview	Interior image generation (4 modes)	6/15 credits
xAI (Grok)	grok-2-aurora	Masked inpainting for region editing, social images	10–20 credits
Kling	kling-v3.0, kling-1.6-pro	Interior + social video generation	15–20 credits

7-Vector Embedding Fusion — The Search Backbone

Every product in the catalog is represented by 7 distinct AI embedding vectors, enabling multi-dimensional similarity search. All stored as halfvec (float16) — 50% storage cost reduction, zero accuracy loss:

Embedding Type	Dimensions	Model	What It Powers
Text	1024D	Voyage AI voyage-3.5	Natural language + keyword search
Visual	768D	SigLIP2	"Find materials that look like this photo"
Understanding	1024D	Voyage AI (from Qwen3-VL JSON analysis)	Spec-based search (dimensions, finishes, surface properties)
Color	768D	SigLIP2	Color palette matching
Texture	768D	SigLIP2	Surface texture similarity
Style	768D	SigLIP2	Design aesthetic matching
Material	768D	SigLIP2	Material type / composition classification

Search results are fused dynamically using 7 query-adaptive weight profiles. The system infers query intent and automatically adjusts how much each embedding type contributes to the final ranking. No competitor in the materials space currently operates at this level of search sophistication.

Complete Feature Set

Feature 1 — Catalog Ingestion (Three Methods)

A. PDF Processing Pipeline (14 Stages)

The primary ingest path. A supplier uploads a product catalog PDF. The platform automatically:

Discovers all products using Claude Haiku (fast boundary detection) + Claude Sonnet (metadata enrichment) — 95%+ detection accuracy
Extracts text with PyMuPDF4LLM (structure-preserving)
Semantic chunking via Anthropic API (800 token max, 100 overlap)
Generates text + understanding embeddings via Voyage AI
Extracts all images from product pages
Runs Qwen3-VL 32B Vision on every image — identifies material type, surface properties, finishes, dimensions visible, color palette
Generates all 5 SigLIP2 visual embedding types per image
Extracts 200+ metadata fields per product (dimensions, material composition, certifications, weight, finish, slip resistance, etc.)
YOLO layout detection + Camelot table extraction on complex page layouts
Creates fully searchable product records with all embeddings linked

PDF Size	Products Extracted	Processing Time
Small (1–20 pages)	1–5	1–2 min
Medium (21–50 pages)	6–15	2–4 min
Large (51–100 pages)	16–30	4–8 min
Extra Large (100+ pages)	30+	8–15 min

9 recovery checkpoints — failure at any stage resumes from the last checkpoint, not from scratch.

B. Web Scraping Automatic product discovery from manufacturer websites via Firecrawl. Discovers product pages, extracts structured data, and runs the same full AI enrichment pipeline.

C. XML Import Structured data import with AI-powered field mapping (Claude Sonnet). Supports recurring scheduled imports, batch processing (10 products at a time, 5 concurrent image downloads), and checkpoint recovery.

Feature 2 — Multi-Modal AI Search

Six search strategies, all available simultaneously with results merged and deduplicated:

Strategy	Mechanism	Best For	Accuracy	Response Time
Semantic	Natural language → text embeddings (1024D)	"sustainable natural stone for wet areas"	85%+	<150ms
Vector	Pure cosine similarity	Precise technical lookups	88%+	<100ms
Multi-Vector (7-fusion)	All 7 embeddings, adaptive weights	Complex multi-attribute queries	90%+	<200ms
Hybrid	Semantic (70%) + PostgreSQL full-text (30%)	Technical product names and codes	90%+	<180ms
Material Property	JSONB filter on 200+ metadata fields	Specification-driven sourcing	95%+	<50ms
Visual / Image Upload	Photo → SigLIP2 embedding → cosine similarity	Inspiration-to-specification workflow	88%+	<150ms
All Combined	All 6 in parallel, merged + ranked	Maximum coverage	Best overall	<800ms

After retrieval, a secondary AI re-ranking pass scores results on visual, understanding, relevance, and material dimensions — surfacing the most contextually relevant products even when they don't exactly match keywords.

Saved Searches with AI Deduplication: Users can save search queries. Claude Haiku automatically detects near-duplicate saved searches (85–95% semantic similarity threshold) and offers to merge them, keeping the search library clean.

Feature 3 — KAI Agent Hub (Three AI Agents)

The Agent Hub (/agent-hub) is the conversational AI interface with memory, multi-turn context, image uploads, and tool-use.

Jarvis — Primary Material Intelligence Agent Model: Claude Sonnet 4.5 | LangGraph + Supabase checkpointer (conversations resume across sessions)

Tools available to all authenticated users:

material_search — 7-vector fusion search across the catalog, with Explainable Search Spec (structured interpretation of query across color/material/style/texture/specification dimensions, displayed as a collapsible card above results)
knowledge_base_search — RAG over the workspace knowledge base
visual_search — automatic image-to-material matching when images are attached
analyze_inspiration_url — Design Inspiration URL Finder: scrape any design URL (Houzz, Pinterest, Dezeen, ArchDaily, manufacturer sites) → extract design tokens (colors, materials, textures, styles) → search catalog for matching products. Accessible via a Globe icon button in the chat toolbar that opens a dedicated modal.

Tools gated to Admin/Owner only:

b2b_manufacturer_search — live web search via Anthropic's built-in web_search_20250305 tool. CEE queries include native-language terms (Polish, Czech, Slovak, etc.)
company_website_scrape — scrapes manufacturer websites via Firecrawl
company_enrichment — enriches company records via Apollo.io
contact_discovery — Hunter.io with Apollo.io fallback
email_validate — ZeroBounce email validation before outreach
save_to_crm — saves enriched company and contact data
seo_keyword_research, seo_article_planner, seo_article_writer, seo_content_analyzer, create_seo_article — full SEO content pipeline
Sub-agents: research_analysis, analytics_analysis, business_analysis, product_analysis

Interior Designer Agent Model: Claude Sonnet 4.5 | Focused on spatial design and visualization

Analyzes room photos: layout, dimensions, natural light, spatial relationships, accessibility compliance
Generates interior design images (text prompt or image reference) — see Feature 4
Matches generated designs back to real catalog materials using 7-vector search
Estimates project costs based on matched materials and dimensions
Generates explorable 3D VR worlds from any design image — see Feature 5
analyze_inspiration_url — paste a design URL to extract materials, colors, and styles and find matching catalog products

Demo Agent Model: Claude Haiku 4.5 | Platform showcases and sales demonstrations. Admin-only.

Feature 4 — Interior Design Generation Suite

The Interior Designer agent generates photorealistic interior design images and videos through four distinct AI generation systems:

A. Replicate Text-to-Image (14 models, run in parallel)

Mode	Models
Text-to-Image	FLUX.1-dev, FLUX.1-schnell, SDXL, SD3, Playground v2.5, Kandinsky 2.2, Proteus v0.2 (7 variations output simultaneously)
Image-to-Image	ComfyUI Interior Remodel, Interiorly Gen1 Dev, Designer Architecture + 4 others

B. Gemini Interior Generation (4 modes)

Mode	Description	Credits
Text-to-Image	Narrative prompt → photorealistic room render	6 (flash) / 15 (4K pro)
Image Edit	Existing room + instruction → transformed image	6 / 15
Floor Plan Render	2D floor plan → photorealistic perspective interior	6 / 15
Floor Plan Diagram	Text description → 2D floor plan layout	6 / 15

The Gemini system uses a two-step style-transfer pipeline — an inspiration image is analyzed by Gemini Vision into a structured design specification, which is then used to edit the target room. This prevents spatial "bleed" (the inspiration image never directly influences the generated geometry) and produces far more accurate style transfers than naive image-to-image approaches.

C. Virtual Staging (with Before/After QA)

Empty room photos → furnished renders via Replicate proplabs/virtual-staging (~56 seconds). 8 room types, 8 furniture styles. 20 credits per run. Accessible as both a standalone tool and a KAI agent tool.

Virtual staging results now include a before/after comparison viewer (VirtualStagingViewer component) with an interactive slider that lets users drag between the original empty room and the staged result. An "Analyze Quality" button triggers a Claude Vision assessment of the staging quality — evaluating lighting consistency, perspective accuracy, furniture scale, material realism, and edge blending — each scored 1–10.

D. Region Editing (Masked Inpainting)

Users paint a zone over any room image using the RegionEditCanvas tool. SAM 2 (Replicate meta/sam-2) generates a pixel-perfect binary mask from drawn hints. Grok Aurora then regenerates only the painted area based on a text prompt (20 credits). For product-specific placement, AnyDoor (ali-vilab/anydoor) places a real product photo into the masked zone with accurate lighting adaptation.

Generation workflow (full):

User describes room style or uploads reference via Interior Designer agent
Platform runs parallel generation (multiple models, multiple variations)
All outputs permanently stored in Supabase Storage
Agent runs 7-vector material matching on generated image — surfacing real catalog products that match
Agent presents cost estimates based on matched materials and room dimensions
User can: generate VR world, create video walkthrough, stage an empty room, or region-edit any element
All generations are credit-billed per operation

E. Interior Video Generation

Any design image can become a video via four AI models:

Model	Credits	Best For
Veo-2 (Google)	30	Cinematic walkthroughs, floor-plan flythroughs
Kling v3.0	20	Product spotlights, before/after reveals, social reels
Wan 2.1 i2v 720p	12	Budget general-purpose
Runway Gen4 Turbo	40	Premium quality

Video type auto-selects the optimal model. Supports 16:9, 9:16, 1:1 aspect ratios. Async polling for long renders (Veo, Runway) with job_id polling pattern.

Feature 5 — VR World Generation (WorldLabs Marble + Spark.js)

Any interior design image — AI-generated or user-uploaded — can be turned into a fully explorable 3D world inside the browser. No app, no plugin, no external tool required.

End-to-end flow:

User clicks "Generate VR" on any design image in the agent chat
Selects model tier (mini for quick preview, plus for high-fidelity final)
Supabase Edge Function uploads the image to WorldLabs Marble API and triggers world generation
Platform polls for completion status, stores all asset URLs in the vr_worlds database table
WorldViewer component renders the 3D Gaussian Splat inline in the chat via Spark.js (Three.js-based, code-split, ~496KB)
The world persists in the database — reloads instantly in future sessions

Model tiers:

Model	Credits	USD Cost	Generation Time	Best For
marble-0.1-mini	50 credits	$0.50	~30–45 seconds	Quick previews, client check-ins
marble-0.1-plus	200 credits	$2.00	~5 minutes	High-fidelity final walkthroughs

Credits are refunded automatically on generation failure.

Output assets per world:

Gaussian Splat SPZ files at 3 quality levels: 100k (draft), 500k (standard), full
Collider GLB (collision mesh for first-person physics)
360° panorama equirectangular image
Thumbnail preview
AI-generated scene caption

Navigation controls:

Mode	Controls
Orbit (default)	Drag to rotate · Scroll to zoom · Right-click drag to pan
Walk (first-person)	WASD to move · Mouse look · Shift for speed boost

Business value of VR worlds: An architect can go from finding a tile in the catalog → generating a room design with that tile → walking through a VR version of that room → requesting a quote — all in one session on one platform. No competitor offers this workflow. VR generation is also a strong upsell moment: mini worlds generate impulse purchases, plus worlds are high-margin repeat purchases for client presentations.

Feature 6 — Moodboards

Visual material collections that serve as both a design tool and a social/portfolio layer.

Private workspace use:

Create named boards and organize materials from search, agents, or the catalog
Add materials directly from search results or agent chat output
Feed a moodboard directly into a Quote (materials pre-populated)
Notes and descriptions per board
Export for client presentations

Public portfolio use (marketplace layer):

Any board can be made public — visible on the user's public profile (/u/:userId)
Public boards appear in the Discover directory for other users to browse
Visitors can comment on public boards (social engagement layer)
One board can be "featured" — pinned to the profile header as a visual portfolio piece with a full-width image preview

As a conversion funnel: Moodboards are a key step in the discovery-to-quote journey. A designer explores materials → builds a board as a design brief → shares it with a client → client requests a quote on specific items from the board → supplier receives a structured enquiry. The board functions as the visual brief.

Feature 7 — Professional Marketplace

A B2B talent and services marketplace built on top of the platform, enabling professionals to be discovered, followed, and hired.

The vision: As the platform grows, it becomes not just a material catalog but the professional network for the design and construction industry. A manufacturer searching for a trusted sourcing agent in a new market, or an architect looking to collaborate with a sustainability consultant, finds and contacts them through the same platform they use daily for material sourcing.

Public Profiles (/u/:userId) — opt-in (is_public = true):

Section	Detail
Identity	Name, company, bio, avatar, location, website, professional type badge
Skills & Expertise	Free-form skill tags (e.g., "Biophilic Design", "LEED Certified", "CEE Material Sourcing")
Services	Rich listings: name, description, price range, previous work portfolio with links
Preferred Factories	Curated list of trusted manufacturers the professional works with
Featured Moodboard	Pinned visual portfolio piece with full image preview
All Public Moodboards	Grid of public boards with previews and comment sections
Social	Follow/unfollow, profile view counter, Hire Me CTA button

Hire Me Flow:

Visitor clicks "Hire Me" or "Hire" on a specific service card
Modal opens with optional service pre-selection
Visitor fills in name, email, message
Request stored in profile_contact_requests table
hire_me_received flow event emitted → triggers configurable automation (email, Slack, CRM)
Platform tracks service interest per professional in analytics

Discover Directory (/discover):

Grid of all public profiles, sorted by profile views
Real-time search: name, company, bio, services, skills, location
Dynamic tag filtering: professional type, services, skill tags
Follow button per card
Follower counts as social proof signal
Up to 60 profiles displayed

Supported professional types: Designer, Interior Designer, Architect, Manufacturer, Brand, Supplier, Sourcing Agent, Consultant

Monetization pathway (current → planned):

Phase	Model
Now	Free — building network density and liquidity
Near-term	Featured listing placements (suppliers/professionals pay to appear higher)
Near-term	Verified badge subscription (trust signal for buyers)
Longer-term	Lead generation packages for suppliers (qualified buyer contact access)
Longer-term	Transaction fee on Hire Me requests (% of project value or flat fee per connected lead)

Feature 8 — Quotes & Project Management

A complete B2B procurement workflow from material selection to project delivery tracking.

Buyer workflow:

Build a quote cart — add products from search, agents, or moodboards
Per item: dimensions (width, height, sqm), quantities, notes
Submit quote request to the supplier
Review and accept/reject each upsell/extra individually
Accept the quote → project timeline auto-initializes
Track 9 predefined milestones from "Quote Accepted" to "Project Completed"

Supplier/Admin workflow:

View all incoming quote requests filtered by custom status tag
Assign color-coded status tags (6 system defaults + unlimited custom)
Attach upsells and extras with pricing
Monitor which extras the buyer accepted/rejected
Update project timeline milestones with notes per step
Manage Upsells library and Timeline Steps configuration

Quote lifecycle: Auto-expire after 30 days of inactivity. Any user action resets the timer. Activity tracking = quote remains active during active projects.

Database: 8 tables — quotes, quote_items, status_tags, upsells, quote_upsells, timeline_steps, quote_timeline, system_settings.

Feature 9 — Price Monitoring

Competitive intelligence for the materials market:

Track competitor prices across multiple sources on configurable schedules (hourly, daily, weekly, on-demand)
Price history database with trend visualization
Configurable price alerts when prices cross defined thresholds
Competitor source management (add/remove monitored sources)
Statistics: min/max/average price trends per product category

Particularly valuable for verified manufacturer/brand users who need competitive pricing intelligence without manual monitoring.

Feature 10 — Knowledge Base & RAG

Document-level intelligence beyond product catalogs:

Upload spec sheets, technical guidelines, research papers, compliance documents
AI semantic chunking with quality scoring (0–1 scale)
Hash-based + semantic deduplication (85–95% threshold)
Chunk relationship mapping (semantic, sequential, hierarchical)
200+ metadata fields extracted per chunk
Queried live by Jarvis agent for context-aware answers with source attribution
Full admin view with quality indicators

Feature 11 — Flow Builder & Automation

Flow Builder — visual no-code automation engine:

Trigger events include: hire_me_received, profile_followed, profile_published, quote_requested, quote_approved, moodboard_shared, moodboard_commented, material_reviewed, user_signup, search_executed, model_3d_created, vr_world_created, and scheduled (cron).

Actions: send email (Amazon SES), Slack notification, call webhook, trigger AI agent, create CRM record.

Background Agent Framework — autonomous long-running AI tasks:

Product Enrichment Agent: continuously enriches catalog products with AI metadata
Material Tagger Agent: auto-tags materials with structured labels
Extensible: new agents implemented via AgentRunner interface
Auto-recovery: tasks stuck >8 minutes re-dispatched automatically (up to 3 retries)
Chain triggers: agents can trigger subsequent agents on completion
Full run history, log viewer, manual trigger from admin panel

Feature 12 — Social Media Suite

AI-powered social media content generation and cross-platform publishing via Late.dev:

Content Generation:

Captions: Claude generates 3 variants per platform (Instagram, Facebook, LinkedIn, TikTok, Pinterest, YouTube, Twitter, Threads), each respecting character limits, hashtag conventions, and platform tone. 2 credits.
Images: Auto-routed by content type — xAI Aurora for lifestyle/people (10cr), Gemini Imagen for product/interior (5cr), FLUX Dev for artistic/textured (6cr). 1:1, 4:5, 9:16, 16:9 aspect ratios.
Videos: Kling 1.6 Pro (15cr, fast) or Veo-2 (30cr, premium) for short-form social videos. Async polling for longer renders.

Publishing (Late.dev):

Users connect Instagram, Facebook, LinkedIn, TikTok, Pinterest, YouTube, Twitter, or Threads via OAuth at their profile settings page
Accounts are per-user (not per workspace — each professional manages their own presence)
Publish immediately or schedule for a future datetime
No credit cost for publishing — uses the workspace's Late.dev subscription
Engagement metrics (likes, comments, shares, reach) synced back to local database via late-analytics
Best posting time recommendations per platform

Business value for the platform: Professionals using Material KAI to design rooms and source materials can publish their work directly to social media without leaving the platform. Every post is a word-of-mouth marketing event for the platform, driving organic discovery. The social suite also creates a natural upsell moment — from free content generation into higher-tier credit packages.

Feature 13 — Email System & Campaigns

Full transactional and campaign email infrastructure via Amazon SES:

Domain verification and management
React Email templates with personalization
Per-recipient tracking: sent, delivered, opened, clicked, bounced
Bounce/complaint handling via SES + SNS webhooks
Campaign system: bulk email campaigns with audience targeting, scheduling, real-time delivery analytics
Admin dashboard at /admin/emails

Feature 14 — CRM (Companies & Contacts)

Built-in CRM for managing supplier and manufacturer relationships:

Company records: name, website, industry, country, Apollo.io enrichment data
Contact records: email (ZeroBounce validated), linked to companies
Records created directly from B2B manufacturer research agent
Full CRUD API: company management, contact management, association management

Feature 15 — Admin Dashboard

Full platform operations suite at /admin:

Section	Purpose
Platform Overview	Cross-workspace analytics, user metrics, hire request funnel, rating distributions
PDF Processing Monitor	Real-time job tracking with 9 checkpoint stages, error logs, retry controls
Analytics	Search patterns, processing stats, AI model usage, cost tracking, agent chat quality ratings
AI Monitoring	Per-model usage, cost breakdown, performance trends, Sentry error integration
Knowledge Base	View/edit chunks, images, products with quality scores
Agent Configurations	Manage AI agent system prompts in real-time (no deployment needed)
Background Agents	Run history, logs, manual trigger, create/edit configurations
User Management	Workspace members, roles, permissions
Quotes	All quote requests with filtering and detail views
Status Tags	Custom color-coded quote status management
Upsells	Upsell library with pricing
Timeline Steps	Project milestone configuration
Email / Campaigns	Email domain management, template library, bulk campaign management
Flows	Visual flow builder
Push Notifications	Workspace-level push notification management

User Segments, Roles & Access Control

Four Functional Groups

Group	Who	Key Privileges
Standard (Buyers)	Designers, Interior Designers, Architects, Sourcing Agents, Consultants, Other	Full search, agents, 3D generation, VR, moodboards, quotes, marketplace
Verified Factory (Suppliers)	Manufacturers, Brands, Suppliers with `factory_verified=true`	All above + Factory Analytics (own data + market trends)
Admin / Owner	`workspace_members.role` = admin or owner	All above + admin panel, B2B tools, SEO pipeline, sub-agents, cross-workspace analytics
Unauthenticated	Not logged in	Public profiles (`/u/:id`) only

Feature Access Matrix

Feature	Standard	Verified Factory	Admin/Owner
7-vector material search	✅	✅	✅
Visual (image) search	✅	✅	✅
Jarvis agent (material + KB)	✅	✅	✅
Interior Designer agent	✅	✅	✅
3D design generation	✅	✅	✅
VR world generation	✅	✅	✅
Moodboards	✅	✅	✅
Quotes & project tracking	✅	✅	✅
Discover directory	✅	✅	✅
Public profile (opt-in)	✅	✅	✅
Factory Analytics — own data	❌	✅	✅
Factory Analytics — market trends	❌	✅	✅
Platform-wide analytics	❌	❌	✅
B2B manufacturer research tools	❌	❌	✅
SEO content pipeline	❌	❌	✅
Sub-agent orchestration	❌	❌	✅
Full admin panel	❌	❌	✅

Business Model

Revenue Streams

1. Credit-Based AI Consumption (Primary) 1 credit = $0.01 USD. All AI operations consume credits deducted from user balance. Purchased via Stripe with volume discounts:

Tier	Spend	Credits per $1	Discount
Standard	$1 – $9.99	100	—
Silver	$10 – $44.99	~111	10%
Gold	$45 – $79.99	125	20%
Platinum	$80+	~143	30%

Volume discounts incentivize larger upfront purchases and reduce churn. A single Stripe product is reused — prices set dynamically at checkout.

2. Supplier / Manufacturer Subscriptions Manufacturers pay a subscription to digitize and publish their catalog on the platform, gaining access to Factory Analytics and appearing in buyer searches. Covers ongoing AI enrichment, updates, and market trend reports.

3. Professional Marketplace — Hire Me Layer Currently free to establish network effects. Planned:

Featured profile/listing placement (pay to appear higher in directory)
Verified badge subscription (trust signal)
Lead generation packages for suppliers (qualified buyer contact access)
Transaction fees on Hire Me connections (% of project value or flat fee per lead)

4. Enterprise / API Access Custom contracts for large manufacturers or distributor networks:

White-label catalog intelligence
API access to 7-vector search against private product database
Private workspace deployments with dedicated AI processing capacity
Custom SLA agreements

AI Cost Structure & Margins

All AI costs tracked in real-time via ai_usage_logs. The platform charges credits at a markup over raw API costs:

AI Operation	Raw API Cost (approx.)	Credits Charged	Platform Margin
Claude Sonnet analysis (1K tokens)	~$0.003–0.015	1–20 credits	20–60%+
Claude Haiku (1K tokens)	~$0.0008–0.004	1–5 credits	50–80%+
GPT-4o-mini (query parsing)	~$0.0002	1 credit	50x+
Voyage AI text embedding (1K tokens)	~$0.00006	Bundled into search	High
Qwen3-VL image analysis (per image)	~$0.02–0.05	2–5 credits	Variable
VR World — mini	$0.50 WorldLabs	50 credits ($0.50)	Pass-through + platform overhead
VR World — plus	$2.00 WorldLabs	200 credits ($2.00)	Pass-through + platform overhead
Interior design (Replicate)	Replicate variable	Credits per image	Variable
Gemini interior (flash)	~$0.02	6 credits	3x markup
Gemini interior (pro/4K)	~$0.06	15 credits	2.5x markup
Virtual staging	~$0.16 Replicate	20 credits ($0.20)	1.25x + overhead
Interior video (Kling v3.0)	~$0.12	20 credits ($0.20)	1.7x markup
Interior video (Veo-2)	~$0.25	30 credits ($0.30)	1.2x + overhead
Region edit (Grok Aurora)	~$0.08 xAI	20 credits ($0.20)	2.5x markup
Social image (Aurora)	~$0.04	10 credits	2.5x markup
Social caption (Claude)	~$0.002	2 credits	10x+ markup
Web scraping (per Firecrawl credit)	~$0.001	~0.1 platform credits	Variable

Blended gross margin target on AI costs: 60–70%. Highest margin on high-frequency, low-cost operations (query parsing, embeddings). Thinner margins on high-cost generation (VR, image gen) where pass-through pricing builds user trust.

Fixed infrastructure costs:

Vercel: Included in Vercel plan at current scale (global CDN, zero egress cost)
Supabase: Managed PostgreSQL, storage, edge function invocations — scales with usage
DigitalOcean: Dedicated server for MIVAA FastAPI backend — predictable fixed monthly cost
HuggingFace Endpoints: Pay-per-use for Qwen3-VL and SigLIP2 inference
Variable: Anthropic, OpenAI, Voyage AI, WorldLabs, Replicate — fully usage-based

Factory Analytics — Supplier Intelligence Dashboard

Verified manufacturers and brands access /factory-analytics:

My Factory Tab (own data):

Product view counts and search appearance frequency
Which products generate the most buyer interest
Quote requests received per product
Hire Me enquiries related to their products
Buyer geographic distribution
Product performance trends over time

Market Trends Tab (platform-wide, anonymized):

Material category demand trends (which types are searched most)
Emerging finish and style trends derived from search query analysis
Seasonal demand patterns
Competitive positioning signals

Platform-Wide Tab (Admin/Owner only):

Cross-factory product performance
User acquisition, activation, retention cohort analysis
Revenue and credit consumption trends
AI model performance and cost efficiency
Search quality metrics

This analytics layer is a significant standalone upsell — providing market intelligence that currently requires expensive industry research reports or trade show attendance.

Production Metrics

Metric	Value
Active users	5,000+
PDFs processed	1,000+
Products cataloged	10,000+
Platform uptime	99.5%+
Search accuracy	85–95%
Product detection accuracy (PDF)	95%+
Material recognition accuracy	90%+
Image classification accuracy	88%+
Search response time	200–800ms
Concurrent query capacity	1,000+/minute
API endpoints	170+
AI models integrated	20+ across 8 providers
Edge functions deployed	60+
Database tables	40+

Technology Differentiation — Why This Is Hard to Replicate

1. 7-Vector Fusion Search Most platforms use one or two embedding types. Fusing 7 specialized vectors with dynamic per-query weight profiles requires custom ML infrastructure, tuned weight profiles per query intent, and deep integration between the vision analysis pipeline and the search layer. This compound in value as more products are added.

2. Understanding Embeddings (Novel Architecture) Qwen3-VL 32B generates structured JSON analysis of every product image — material properties, surface characteristics, visible dimensions, finish type. This JSON is then embedded via Voyage AI into a 1024D "understanding" vector. This enables queries like "find matte surfaces with slight veining, suitable for wet external areas" — which no traditional image or keyword search can handle. This is a proprietary architecture not seen replicated elsewhere.

3. Full Vertical Integration Ingestion + enrichment + search + design generation + VR visualization + quote management + marketplace in one product. Each layer creates switching costs. A user with their catalog ingested, moodboards saved, and project quotes tracked is not going to migrate to a competitor easily.

4. 14-Stage Pipeline with Checkpoint Recovery Enterprise-grade PDF processing at 95%+ accuracy with YOLO layout detection, table extraction, and 9-stage checkpoint recovery. This took significant engineering and handles everything from scanned PDFs to complex multi-product catalogs.

5. Halfvec Storage (50% Infrastructure Cost) All 7 embedding types stored as float16 via pgvector. Halves vector storage costs with zero accuracy loss — critical for scaling to millions of products.

6. Real-Time Automation + Background Agents Event-driven flows and autonomous background agents keep improving product data and automating connections without human intervention. Compounds value over time.

7. Social-to-Platform Flywheel Every interior design created on the platform can be published directly to Instagram, LinkedIn, TikTok, etc. via the built-in social media suite. This turns users into organic distributors of platform-created content, each post pointing back to a professional profile on materialshub.gr. No competitor offers this closed loop from design → social publication.

8. Full Vertical Integration of the Design Workflow No competitor currently offers: ingestion → AI search → agent interaction → 3D generation → virtual staging → region editing → interior video → VR walkthrough → social publication → quote → project tracking — all in one product. Each feature layer increases switching costs. A user who has cataloged products, built moodboards, generated designs, and tracked a project on this platform has strong reasons to stay.

Competitive Landscape

Platform	Category	Gap We Fill
Architonic / Archiproducts	Manual product directory	AI ingestion (zero manual effort), 7-vector search, agent interactions, buyer tools
Material Bank	Physical sample delivery	Fully digital, globally accessible, no logistics dependency
Trade Shows (Coverings, Salone del Mobile)	Periodic offline discovery	Always-on, searchable, data-driven
Manufacturer websites	Brand-specific catalogs	Cross-manufacturer search, buyer-side design and procurement tools
Generic AI chatbots (ChatGPT, etc.)	General Q&A	Deep material domain knowledge, live catalog integration, visual search, full procurement workflow
BIM tools (Revit, ArchiCAD)	Design software	Complementary discovery layer — not competing
Spec platforms (NBS, Arcom)	Technical specification databases	Richer visual and AI-assisted discovery, broader product coverage, integrated procurement

The defensible position: No competitor currently offers the combination of auto-ingestion + multi-modal AI search + AI design generation + VR visualization + marketplace in one product. Each element alone is replicable; the vertical integration combined with 7-vector search and the understanding embedding architecture creates a meaningful moat.

Roadmap

Live in Production (March 2026)

7-vector fusion search with query-adaptive weight profiles
Understanding embeddings (spec-based search via Qwen3-VL + Voyage AI)
AI search re-ranking (Claude post-retrieval re-ordering with explanations)
PDF ingestion pipeline (14 stages, 9 checkpoints, YOLO layout detection)
Web scraping ingestion (Firecrawl)
XML import with AI field mapping
KAI unified agent (material search, KB RAG, visual search, B2B research, SEO)
Interior Designer agent (room analysis, 3D generation, material matching, cost estimation)
14 Replicate generation models (7 text-to-image, 7 image-to-image)
Gemini interior generation (4 modes: text-to-image, image-edit, floor-plan-render, floor-plan-text)
Virtual staging (AI-furnished rooms from empty photos — Replicate proplabs, 20cr)
Region editing / masked inpainting (SAM 2 + Grok Aurora, 20cr)
Interior video generation (4 models: Veo-2 30cr, Kling v3.0 20cr, Wan 2.1 12cr, Runway 40cr)
Social media suite (caption/image/video generation + Late.dev publishing across 8 platforms)
VR world generation (WorldLabs Marble + Spark.js, orbit + first-person walk navigation)
Professional marketplace (profiles, services, Hire Me, Discover directory)
Moodboards (private + public, comments, featured board)
Quotes system (buyer + supplier workflow, upsells, project timeline)
Price monitoring (multi-source, alerts, trends)
Knowledge Base & RAG
Flow builder & automation engine (cron, webhook, event triggers; SMS/email/HTTP actions)
Background agent framework (6 types: product enrichment, material tagging, social analytics, factory enrichment + auto-recovery)
Factory Analytics dashboard (own data + market trends + platform-wide for admins)
Full admin dashboard (analytics, monitoring, prompt management, CRM, campaigns, background agents, flows)
Email system (Amazon SES, React Email, campaign management)
Billing & credits (Stripe subscriptions + credit packages, customer portal)
Duplicate detection & product merging (factory-scoped, undo capability)
Collaborative filtering recommendations
CRM (contacts, companies, users with role management)

Near-Term

Sourcing Agent elevated B2B tools (without requiring full admin role)
Credit quota differentiation by user tier
Verified factory elevated KAI agent tools
User onboarding flow with professional type self-selection
Hire Me transaction fee monetization layer
Featured listing placements in Discover directory (paid)

Strategic / Long-Term

Mobile app — point camera at a material → instant catalog match
White-label API for enterprise manufacturers (catalog intelligence as a service)
Multi-language search and localization (CEE dual-language B2B queries partially implemented)
Full marketplace transaction layer: RFQ → Quote → Purchase Order → Delivery tracking
Verified badge subscription for professionals and suppliers
Third-party integrations: Revit plugin, ArchiCAD plugin (material spec directly to BIM models)
Carbon footprint and sustainability scoring per product (emerging EU regulatory requirement)

Platform & Team

Platform: materialshub.gr API: v1api.materialshub.gr/docs Repository: github.com/creativeghq/material-kai-vision-platform Status: Production — 5,000+ active users, 99.5%+ uptime Version: 3.5.0 — March 2026

This document is confidential and intended for investors, strategic partners, and due diligence purposes only. All metrics are as of March 2026.