Artrelo

Property maintenance coordination is slow, manual, and inconsistent. Requests arrive as free-form text and photos, and someone has to judge what the issue is, how urgent it is, whether it's a safety hazard, and which vendor should handle it. Done by hand this is error-prone — and the failure mode isn't just a slow ticket, it's a life-critical hazard being mis-triaged.

Property managers and their dispatch teams, who need fast and consistent triage, and tenants, who report issues (often with a photo) and need urgent problems handled reliably.

Artrelo runs each incoming request through a multi-stage LLM pipeline that classifies the incident, asks de-duplicated follow-up questions, and synthesizes a structured incident record. Independently of the model, a deterministic safety layer screens for life-critical hazards, and a routing-readiness engine decides whether a case is safe to auto-dispatch or should be escalated to a human.

1. Step 1IntakeFree-form request + photo from the tenant enters the pipeline.
2. Step 2LLM triageStructured outputs classify the incident and generate de-duplicated follow-up questions.
3. Step 3Incident recordA synthesized, structured record of the issue and its evidence.
4. Step 4Routing-readinessA six-state engine gates dispatch on confidence, evidence, cost tier, and responsibility.
5. Step 5Dispatch / reviewReady cases auto-dispatch to a vendor; uncertain or high-risk cases route to a human.

• Structured LLM outputs

  Triage uses constrained, structured outputs rather than free text, so downstream logic can depend on parseable, well-typed fields instead of prose.

• Safety independent of the model

  A deterministic safety layer runs separately from the LLM to detect life-critical hazards and force-escalate emergencies, so safety never depends on model behavior.

• A readiness state machine, not a threshold

  Rather than a single confidence cutoff, a six-state routing-readiness engine gates automated dispatch on AI confidence, evidence completeness, cost tier, and tenant responsibility.

• Graceful degradation

  Deterministic fallback logic keeps the pipeline functioning when the model fails, degrading to safe defaults instead of breaking.

• Determinism over flexibility for safety

  A rule-based safety layer is less adaptive than an LLM, but it is auditable and predictable — the right trade for life-critical decisions.

• Friction for uncertain cases

  Gating dispatch on readiness adds a human step for ambiguous cases, trading throughput for a lower chance of an unsafe automated action.

• Constrained outputs over open generation

  Structured outputs limit what the model can express, in exchange for reliability and easier integration with deterministic logic.

• A safety layer independent of the LLM detects life-critical hazards and force-escalates emergencies.
• Uncertain or high-risk cases are routed to human review rather than auto-dispatched.
• Deterministic fallback logic provides graceful degradation during model failures.
• Hazard-specific guidance is returned deterministically, not left to model discretion.

Live at artrelo.com and in active development. Capabilities described here reflect the system's design; this writeup avoids quoting performance numbers that aren't independently verified.

• Broaden deterministic hazard coverage and hazard-specific guidance.
• Richer vendor matching and cost-aware routing.
• Feedback loops that use dispatch outcomes to improve triage quality.
• Observability on triage decisions and routing-readiness transitions.