Technology · For CTOs and IT directors
How the workforce actually works.
Four-memory cognitive loop. Knowledge graph instead of vector chunks. Computer Use + MCP for software integration. Tenant-resident execution. The technical answers your security questionnaire needs.
The user surface
All of this, from a text box.
The architecture below — four memories, MCP, Computer Use, polycentric orchestration — runs in response to a user typing a brief. No new IDE. No prompt-engineering ritual. Plain language, in.
Scroll to advance
The user surface
All of this, from a text box.
Scroll to advance
Cognitive architecture · 4 memories
Four memories. One cognitive loop.
Most AI products do vector RAG and call it intelligence. Construction projects have dates, authors, relationships, procedural memory. We model all four — and every agent reasons across all four on every call.
Step 0 · Before the cognitive loop
Ingestion is automatic. You don't upload.
Point the agent at your existing SharePoint / OneDrive / DMS folder once. From that moment on, every new document, every email, every revision is indexed into the four memories as it arrives. Permissions inherited. No upload step. No duplication. The four memories below populate themselves.
47 documents indexed
Ingestion processing 47 documents
Memory 1 · RAG
What documents say.
Semantic vector search across 10,000+ pages. "Find every force-majeure clause" — answered in seconds, sourced to document + page. Qdrant under the hood.
Memory 2 · Graph
What's connected to what.
Neo4j knowledge graph. 21 organic base labels, 42+ relationship types. Document-as-graph: a 60-page contract becomes 200+ structured nodes. Variation → invoice → clause → penalty.
Memory 3 · Timeline
Who did what when.
PostgreSQL append-only event log. Every document version, every approval, every email, timestamped to the second. Immutable. The legal-grade chronology a dispute needs.
Memory 4 · Extensions
How the firm works.
Procedural memory. Reusable scripts, encoded protocols, normative calculations. When a senior leaves, their methods stay. The 5th dimension other AI products don't model.
Cognitive loop: CONTEXT → SEARCH → NAVIGATE → READ → REASON → ACT → ENRICH (loops as needed). Every agent. Every call. Not a linear pipeline.
Memory 3 in the app · Timeline / event stream
What the append-only event log looks like.
Every document version, every approval, every email, every agent action — timestamped to the second, immutable, hashable. The legal-grade chronology a dispute needs, ready before the dispute arrives.
Event stream
Time-bar at risk: Avenant #3 — 7 days remaining
Legal agent·May 4 · 09:14
Persisted across the Vonbuild knowledge stack
Project timelineDBProject knowledgeRAGCompany knowledgeRAGProject mappingGraphCompany mappingGraph
Email sent: relance livraison acier
Operations agent·May 4 · 08:42
Persisted across the Vonbuild knowledge stack
Project timelineDBProject knowledgeRAGCompany knowledgeRAGProject mappingGraphCompany mappingGraph
OneDrive indexed 47 documents
System·May 3 · 17:30
Persisted across the Vonbuild knowledge stack
Project timelineDBProject knowledgeRAGCompany knowledgeRAGProject mappingGraphCompany mappingGraph
Confirmation rejected: send claim letter
Kais Shili·May 3 · 14:08
Persisted across the Vonbuild knowledge stack
Project timelineDBProject knowledgeRAGCompany knowledgeRAGProject mappingGraphCompany mappingGraph
Document-as-graph · The differentiator
A contract is 200 nodes, not 1 file.
Other AI products treat a document as a chunked blob. We explode it into a structured node tree, then weave the graph across documents.
7 document patterns shipped
Contract
Section → Article → Clause → Penalty → Notification window
BOQ
Lot → SubLot → LineItem → Rate → Quantity
Invoice
InvoiceLine → CLAIMS_WORK → LineItem (BOQ cross-ref)
Meeting
Decision → Person → Action → DueDate
Tender
Requirement → ScoringCriterion → MandatoryClause
Drawing
Drawing → Revision → SUPERSEDES → PriorDrawing
Email
Email → From → To → Mentions → Attachment → ThreadOf
Outcome: the graph holds 30–500 nodes per indexed document instead of 1 generic Document node. Cross-document relationships are queryable in milliseconds. Variation #3 ↔ Invoice #12 ↔ Delay event #47 ↔ Engineer's drawing rev_3.2 — surfaced as a single subgraph.
Execution model · 4 routes
How agents actually do the work.
Every agent task tries these four routes in order. Whatever produces a real verifiable result, ships. None of the four works → reported back, never fabricated.
01
A tool that already exists
Reach for what's on the shelf.
Your firm has built this workflow before. Or Vonbuild ships it out of the box. The agent picks the right one and runs it — like your engineer reaching for a checklist they trust.
02
A tool the agent builds
Make the missing one.
Need a calculation nobody's done in your firm yet? Hand the agent the norm — EN 13474 for glazing, Eurocode 2 punching shear, RT2020 thermal. It builds the calculator, checks it against known cases, runs your project. Next project, that calculator's already in your firm's library.
03
A tool that talks back
Connect straight through.
Some of your software — Revit, a growing list of cloud tools — lets the agent connect directly, without going through screens. No clicking, no waiting. The cleanest route when it's available.
04
A tool the agent uses like you would
Click through, like an engineer would.
Everything else — Robot, ETABS, Coupa, Procore, your supplier portals. The agent logs in with your licence and works the interface exactly as your engineer or QS would. Same software, same result, same paper trail. None of your hours.
Or — you hear back, with the reason
No made-up numbers
When none of the four routes works, the agent tells you what it tried, where it stopped, and what would unblock it. It does not invent a number to fill the gap. The output on your screen is always a real result from a real tool — or there's no output.
Model Context Protocol
MCP · the standard for tool access.
Any software with an MCP server connects natively. Revit 2027, Outlook, Teams, SharePoint, internal APIs. No bespoke integration project — just wire the MCP server.
What MCP gives you
Authenticated, audited, programmatic.
No GUI scraping. No brittle browser automation when there's a real API. Authentication respects your existing identity gates. Every MCP call logged in the audit trail.
Shipped MCP tools
20+ tools today.
Microsoft Graph (Outlook / SharePoint / Teams), Google Workspace, Qdrant, Neo4j, PostgreSQL, schedule(), graph_enrich_node(), document hashers. Plus the Vonbuild MCP sidecar exposed to agents.
Custom MCP (Enterprise)
Wire your in-house systems.
SAP, Navision, custom ERP, internal cost databases — our architects build the MCP server, your IT team approves the access scopes. Per-tool ACLs, audited per call.
plugin.mcp.json · sample
{
"mcp_servers": {
"sap-erp": {
"url": "https://mcp.firm.internal/sap",
"auth": "oauth2_passthrough",
"scopes": ["read:purchase-orders", "read:invoices", "write:purchase-orders"],
"audit_level": "every_call",
"access_roles": ["procurement-manager", "finance-controller"]
},
"revit-2027": {
"url": "localhost:7117",
"auth": "user_session",
"scopes": ["read:model", "read:elements", "read:parameters"],
"audit_level": "every_call"
}
}
}Marketplace · Internal extensions
Procedural memory that compounds.
Reusable scripts · workflows · normative routines · knowledge packs · always-on agents. Authored by your team, governed by your DSI.
Scope hierarchy
Project → tenant promotion.
Extensions start at project scope. Author requests promotion. DSI reviews (security check, IP check, scope confirmation). Approved extensions become tenant-wide.
Auto-coded mini-apps
Agent writes the calculator.
When no extension exists, the agent writes the Python script from the normative reference. Sandboxed run. Validated against reference cases. Promotable after DSI approval.
Auto-assignment
Rules, not configuration.
Extensions self-assign by project metadata: jurisdiction, project phase, contract type, building typology. New project → relevant extensions pre-loaded. No setup project.
Five extension types
workflow · multi-step orchestrated
knowledge · curated reference pack
mini-app · focused tool
agent · always-on triggered
process · firm-encoded procedure
Deployment architecture
Tenant-resident. Identity-inherited. Multi-posture.
Production runs on bare-metal in Germany. K3s + Cilium + GitOps via ArgoCD. Per-tenant Kubernetes isolation. NetworkPolicies between tenants.
Compute · EU bare-metal
- • Hetzner FSN1 · Auction-tier dedicated servers
- • AMD Ryzen 9 / ECC RAM / NVMe RAID
- • No US hyperscaler dependency for storage
- • Per-pod isolation, NetworkPolicies enforced via Cilium eBPF
Data layer · split by concern
- • PostgreSQL — timeline events, audit log, transactional state
- • Neo4j — knowledge graph (21 base labels, 42+ relationships, organic extension)
- • Qdrant — semantic vector store for RAG
- • JuiceFS — workspace files, model artefacts
Identity · M365 / Workspace inheritance
- • Keycloak OAuth2 / OIDC for identity broker
- • SAML SSO from your tenant
- • SCIM provisioning supported
- • DLP labels respected · retention policies inherited
LLM layer · LiteLLM proxy
- • Anthropic Claude · OpenAI · Gemini · enterprise endpoints
- • Zero-retention on enterprise tiers
- • Token budgeting + per-tenant rate limiting
- • Bring Your Own LLM at Enterprise tier · any provider, your keys
Full architecture diagrams + threat model + CAIQ + SIG questionnaire available under NDA at the technical deep dive.
Scheduler · Event-driven runtime
Agents schedule their own future.
Email-driven · document-driven · time-driven. Tasks queued, surfaced to your Outlook "Vonbuild" calendar, auto-cancelled when resolved.
Triggers
Email arrives in matter folder · file lands in OneDrive · date threshold crossed · upstream agent enriches the graph with a flagged node.
Frequencies
One-shot · daily · weekly · biweekly · monthly · cron expressions. Timezone-aware (DST respected).
Calendar surface
Scheduled tasks pushed via Microsoft Graph to a dedicated "Vonbuild" calendar in user's Outlook. Read-only in Outlook (modifications via Vonbuild UI). Cancellation deletes the event.
Auto-cancel logic
An incoming email that resolves a scheduled task (e.g. the PPSPS document arrives) auto-cancels the task. Agent decides via prompt context, scheduler removes from heap + deletes the Outlook event.
Implementation: PRD 107 · scheduler worker pod with PG LISTEN + in-memory heap. Currently in beta — stability improvements rolling out through Q3 2026.
Live demos · The two scheduling surfaces
Same engine — your team schedules explicitly, the agent schedules autonomously.
Your users open these modals to schedule a one-shot follow-up or a recurring sweep. The agent uses the same underlying scheduler service — no UI — to set its own reminders, re-runs, and triggers. Both populate the Outlook "Vonbuild" calendar.
Option A · One-time
Option B · Recurring
Integrations
Operates inside your existing stack.
No migration. No new UI to roll out. The agent navigates your software the way your team does — via MCP where available, browser elsewhere.
Microsoft 365 · native
OutlookSharePointTeamsOneDriveCalendar
Via Microsoft Graph API. OAuth2 passthrough with your tenant. SCIM provisioning. Read+write per scope.
Professional software · MCP or browser
Revit 2027 (native MCP)Robot StructuralETABSSAP2000Coupa · AribaProcore · Plangrid · DaluxP6 · MS Project · Asta
Anything with an MCP server connects natively. Anything browser-accessible drivable via Computer Use (Playwright shared session, your authenticated login).
Internal systems · Custom MCP (Enterprise)
SAPNavisionIn-house ERPCustom cost DBLegacy AS/400
Our architects build the MCP server. Your IT defines scopes. Per-call audit. Enterprise tier.
60 minutes. Your security team. Our architects.
Bring the CAIQ / SIG questionnaire. Bring your CISO's hardest question. We come with architecture diagrams, threat model, audit-log samples, MCP scopes, deployment topology. No marketing slides.