Knowledge was separated from work
SOPs, policies, training material, and business rules were disconnected from the operational system of record.
FIELD NOTES / REAL ESTATE & PROPERTY OPERATIONS
Property operations.
AI-assisted. Human-approved.
Rent Manager AI was designed as a governed intelligence layer for property teams—grounding answers in company knowledge and operational data, then routing consequential actions through explicit approval, verified execution, and audit.
REAL ESTATE CASE STUDY
Rent Manager AI
Property operations.
Governed AI workflows.
Rent Manager AI / Operations workspacePRODUCT EXPERIENCE
GROUNDED OPERATIONSAsk in plain language. Keep the controls.
Operational data, company knowledge, and permitted tools come together without giving the model unchecked authority.
Follow the interaction ↗UnderstandRetrieveProposeApprove & verify
DOMAIN
Property management operations
WHAT WE DESIGNED
Governed AI operations layer
OUR ROLE
Product, AI & integration architecture
01 / THE OPERATING CHALLENGE
The answer exists.
The context is scattered.
Property teams work across leases, ledgers, maintenance records, tasks, tenant communication, SOPs, and policy documents. The question is simple; assembling a reliable answer often is not.
An AI interface can reduce that friction—but only if it respects permissions, separates advice from execution, and never reports success before the property system confirms the change.
The engagement defined an architecture where the model can understand, retrieve, reason, and propose. Application logic remains responsible for authorization, approval, execution, verification, and audit.
Answers crossed too many surfaces
Receivables, occupancy, maintenance, tasks, and documents required different views and repeated reconstruction.
Agentic writes needed a hard boundary
Charges, renter data, lease records, and work orders could not be changed by an unrestricted model.
02 / INSIDE THE PRODUCT DESIGN
One conversational layer.
Explicit operational control.
Explore the supplied product concepts for grounded questions, approval-gated work, maintenance visibility, and portfolio context.
Ask the business question—not the database question.
The workspace translates a property-operations request into controlled tool use and presents a grounded, readable response with its supporting context.
THE GOVERNED AI LIFECYCLE
Understand→Retrieve→Reason→Propose→Approve→Execute→Verify→Record
03 / THE APPROVAL BOUNDARY
Fast on the read.
Deliberate on the write.
Read-only questions can use permitted operational tools and relevant company knowledge. A consequential change becomes a reviewable proposal—showing the action, current state, intended state, policy context, and risk—before any write is attempted.
Look under the hood ↘01
Ground the request
Tenant-scoped knowledge · role permissions · current operational data
Return a grounded answer
Use permitted tools, expose supporting context, and preserve traceability.
Create an approval proposal
Show before/after state and do not write until an authorized person approves.
02
Execute through the integration layer
Application logic authorizes and invokes the approved system operation.
03
Verify before claiming success
Confirm the downstream result, surface failure honestly, and write the audit record.
Interpret intent
Turn a natural-language request into an explicit operational need.
Retrieve context
Use tenant-safe documents, policies, permissions, and live system data.
Approve change
Keep sensitive write operations behind a human authorization gate.
Verify & record
Confirm downstream state and preserve a traceable execution history.
04 / THE TECHNOLOGY & THE REASON
An AI experience.
An application-controlled system.
The proposed architecture separates conversation, retrieval, orchestration, approval, integration, and observability—so model reasoning never becomes an implicit authorization layer.
OPERATOR EXPERIENCE
Next.js + TanStack Query
Chat, knowledge, approvals, and usage views share a permission-aware product surface.
Why this layer?
Operational state stays explicit while server data, pending actions, and verification states remain easy to reason about.
APPLICATION CONTROL
FastAPI + Pydantic
A typed modular monolith owns authorization, workflows, validation, and controlled business execution.
Why not let the model execute?
Business rules and permissions belong in deterministic application code, not probabilistic model output.
MODEL LAYER
AWS Bedrock + Titan
A gateway coordinates model access and embeddings while keeping providers behind an application boundary.
Why a model gateway?
Central control improves configuration, observability, and future model portability.
TENANT-SAFE KNOWLEDGE
PostgreSQL + pgvector
Operational metadata and vector retrieval share a relational core designed around organization isolation.
Why tenant-aware retrieval?
Document relevance is useful only when every query remains inside the correct organization boundary.
SYSTEM INTEGRATION
MCP + approval service
Read and write tools are exposed through a controlled integration layer, with writes routed through approval.
Why isolate the connector?
System credentials, tool permissions, and downstream behavior can evolve without weakening the application boundary.
OBSERVABILITY & PLATFORM
Langfuse + Redis + ECS
Tracing, token and cost visibility, operational support, and container deployment were designed into the platform.
Why observe AI separately?
Teams need model, latency, token, cost, and organization attribution to govern a real service.
The architecture visual and product screens reproduce the supplied engagement materials. Screen names and values are illustrative product-design data, not measured client outcomes.
Explore the complete specified stack
The system specification covers experience, application services, retrieval, model access, integration, security, observability, and cloud delivery.
- Experience
- Next.js · Tailwind CSS · CSS variables · shadcn/ui · TanStack Query
- Application
- Python · FastAPI · Pydantic · SQLAlchemy · Alembic · modular monolith
- Data & retrieval
- PostgreSQL · pgvector · Redis · Amazon S3 · organization-aware design
- AI
- AWS Bedrock · model gateway · Titan Embeddings · controlled orchestrator · Langfuse
- Integration & security
- Rent Manager MCP · RBAC · server-side authorization · credential references · audit trail
- Cloud delivery
- AWS ECS Fargate · VPC · IAM · Secrets Manager · Docker · GitHub Actions
Technology and platform names identify the specified solution, not partnerships or endorsements.
A CONVERSATION ABOUT YOUR PROPERTY OPERATION
Which workflow still sits
between your PMS, inbox, and SOPs?
Bring one leasing, receivables, maintenance, tenant-service, or portfolio workflow. In 20 minutes, we’ll map the useful AI boundary, the approvals that matter, and a practical engineering next step.
20 minutes · Your workflow, control points, and next step
20 minutes · Your workflow, control points, and next step