skip to content
home / projects / buildingrelay

BuildingRelay

AI-assisted communication and case management for building services — live, staged, and working.

BuildingRelay is a communication and case-management platform purpose-built for the messy reality of building service work. Tenants, property managers, service providers, and vendors communicate across email, SMS, and phone calls — fragmented channels with no shared history. BuildingRelay unifies intake through email and SMS, uses a dual-path CaseAgent (LLM reasoning with deterministic fallback) to route, triage, and respond, and keeps a permanent searchable case history. It has a complete React frontend, a media pipeline with signed URLs and thumbnails, a building-memory system that extracts knowledge and detects recurring issues, an approvals workflow, and a safety ops cockpit — all built with aggressive safety defaults: every external action — sending an email, closing a case, dispatching work — requires human approval.

Live staging · AI case agent · dual-path LLM · safety-first defaults case study
PythonFastAPIPostgreSQLReactSendGridIMAPGmail APIDockerDigitalOcean

01. The problem: fragmented communication in building services

Building service work generates a firehose of communication across email, text messages, photos, and phone calls. Tenants report issues in whatever channel is fastest. Property managers forward emails. Vendors send photos and invoices. Nothing is threaded, nothing is searchable, and the same issue gets reported multiple times through different channels. Existing property-management software is designed for rent collection and lease tracking, not for the operational chaos of maintenance coordination. The goal was not to force everyone into a new app — it was to meet people in the channels they already use and bring coherence to the conversation from there.

02. The CaseAgent: dual-path reasoning with safety defaults

Every incoming message hits the CaseAgent — a dual-path system that runs LLM reasoning and deterministic rules in parallel. The LLM path classifies intent, extracts entities, and suggests actions. The deterministic path validates against known patterns, checks safety rules, and enforces policy constraints. If the two paths agree, the action proceeds. If they diverge, the system defaults to the safer path and flags for human review. The CaseAgent handles communication routing and response drafting, but no action with an external effect — sending an email, closing a case, dispatching high-risk work — executes without explicit human approval. This dual-path architecture means the system is flexible enough to handle novel situations but conservative enough to never act on a hallucination.

03. Email and SMS intake: meeting users in their channels

The intake layer supports IMAP, Gmail API, SendGrid, and Mailgun for email, plus SMS gateway integration. Incoming messages are parsed, deduplicated, threaded into cases, and routed to the CaseAgent. The system handles forwarded threads, photo attachments, and multi-recipient conversations. Outbound communication goes through SendGrid (email) and the SMS gateway, with all communication recorded in the case history. The person on the other end never needs to log into a portal — they just reply to the email or text, and the system handles the rest.

04. AI-powered communication adaptation

Not every message needs the same tone or language. BuildingRelay includes a communication adaptation layer that can translate messages between languages and adjust tone — formal for compliance-sensitive communications, casual for tenant-facing messages, technical for vendor coordination. The adaptation is configurable per-property and per-channel, so a single platform can serve residential buildings, commercial properties, and industrial sites with different communication norms. All adaptations are logged and auditable.

05. Building memory: knowledge extraction and recurrence detection

Over time, the same issues recur — the same boiler fails in the same building, the same tenant reports the same leak. BuildingRelay's memory system extracts structured knowledge from case histories: which units have chronic issues, which vendors resolve problems fastest, which seasons trigger which failure modes. A recurrence detection engine identifies patterns across cases and can proactively suggest preventive maintenance. The memory is stored per-building so knowledge accumulates locally — a building's history informs its future without cross-contaminating across properties.

06. Media pipeline, approvals, and the safety cockpit

Photo evidence is critical in building services — a picture of a leak tells a very different story than a text description. The media pipeline handles uploads, generates signed URLs with expiration, creates thumbnails, and attaches photos to the correct case. The approvals workflow routes repair quotes, vendor assignments, and emergency dispatches through the right chain — configurable per property and per dollar threshold. The safety ops cockpit gives property managers a real-time view of open safety issues, response times, and escalation status. 67 test files cover auth, case agent, communication, email, knowledge, media, people, safety, and worker subsystems — the test suite is a first-class artifact, not an afterthought.

◆ Outcome

BuildingRelay is live as a staging environment at buildingrelay.theducklabs.com with a complete backend (FastAPI, PostgreSQL) and a redesigned React frontend. It handles email and SMS intake, routes cases through a dual-path CaseAgent with LLM and deterministic reasoning, adapts communication by language and tone, extracts building memory across cases, and enforces safety-first defaults. 67 test files across every subsystem. This is not a demo — it is a deployed, staged platform solving a real communication fragmentation problem in an industry that runs on email forwards and sticky notes.