Attune
Vienna, Virginia
IoT-enabled PropTech / Building Intelligence — Environmental Monitoring
Amazon Bedrock Anthropic Claude Sonnet 4 AWS EC2 AWS IAM AWS S3 Amazon RDS React Chain-of-Thought (CoT) Prompting Structured JSON Prompt Templates GitHub
Attune
Vienna, Virginia
IoT-enabled PropTech / Building Intelligence — Environmental Monitoring
Amazon Bedrock Anthropic Claude Sonnet 4 AWS EC2 AWS IAM AWS S3 Amazon RDS React Chain-of-Thought (CoT) Prompting Structured JSON Prompt Templates GitHub
Attune is a leader in IoT-enabled PropTech, delivering real-time environmental intelligence across schools, hospitals, and commercial buildings through a unified, cloud-hosted platform. While Attune’s sensor network generated rich, continuous data across multiple building systems, its existing rules-based alerting couldn’t explain why anomalies occurred, leaving facility managers to manually interpret complex, multi-sensor events. Avahi designed and delivered a GenAI-powered anomaly interpretation system on Amazon Bedrock, using Anthropic’s Claude Sonnet 4 to generate natural language root-cause explanations, ranked causes, and actionable guidance directly from live sensor data. The result: a 19% reduction in triage time and a scalable AI foundation ready for enterprise-wide deployment.
Attune is a PropTech innovator with 45+ patents in integrated IoT solutions, providing building owners, operators, and facility managers with real-time visibility into indoor air quality (IAQ), HVAC performance, energy usage, and water risk. Their cloud-hosted platform aggregates sensor data across diverse, mission-critical environments, including K–12 schools and hospital systems, empowering operators to optimize building performance, ensure regulatory compliance, and reduce operational costs at scale.
Attune’s platform was already capturing a continuous stream of high-fidelity sensor data, CO₂, temperature, humidity, occupancy, particulate matter (PM2.5, PM10), TVOC, and water flow, across dozens of customer sites. When sensor thresholds were breached, the system surfaced “Action Cards”: predefined alerts that told operators what had triggered, but not why it happened or what to do about it in context. These static, pre-written responses couldn’t account for the complexity of real-world building systems, where a single anomaly may involve cascading interactions between HVAC units, ventilation zones, occupancy patterns, and environmental conditions.
For facility managers (many of them non-technical) this created a significant operational burden. Interpreting high-dimensional sensor data across interconnected building systems required expertise and time that most teams simply didn’t have. Alerts went under-investigated, root causes went unresolved, and the risk of SLA breaches, energy waste, and compliance gaps grew with every undiagnosed anomaly.
The consequences of inaction were clear: energy inefficiencies, undetected ventilation failures, and water anomalies would persist longer, increasing operational costs and exposing building occupants, including children and hospital patients, to avoidable environmental health risks. Attune needed AI that could do more than flag problems. They needed a system that could reason through them.
Attune’s existing infrastructure was already built on AWS, making Amazon Web Services the natural platform for extending their capabilities with generative AI. Amazon Bedrock provided a fully managed, enterprise-ready foundation for LLM inference, eliminating the need to manage model infrastructure while offering access to leading foundation models with the security and compliance controls required for Attune’s regulated customer environments. AWS IAM enabled precise, role-based access controls within Attune’s own AWS account, and the broader AWS ecosystem — including EC2, S3, and RDS, allowed the solution to integrate directly with Attune’s existing data storage and compute infrastructure without architectural disruption.
Avahi brought a structured, evidence-driven approach to AI that aligned with Attune’s need for reliability in mission-critical environments. Rather than moving straight to development, Avahi proposed a formal Discovery Assessment, a rigorous evaluation phase that included empirical testing of three AI models against real Attune sensor data, a data quality report, and a quantified business case before a single line of production code was written. This discipline gave Attune confidence that the solution would be grounded in their actual operational reality, not generic AI assumptions.
As a premier-tier AWS partner, Avahi had the technical depth to navigate complex Bedrock configurations, IAM trust boundaries, and quota management — challenges that surfaced during the project and were resolved without delay. Avahi also demonstrated the flexibility and accountability that Attune needed: when early feedback indicated the initial approach wasn’t delivering sufficient differentiation over the existing Action Cards, the team rearchitected the solution in response, a pivot that ultimately produced the strongest version of the product and earned formal client sign-off.
Avahi delivered a GenAI-powered IoT anomaly interpretation system built on Amazon Bedrock, integrating Anthropic’s Claude Sonnet 4 directly into Attune’s sensor data pipeline. The solution was developed across a structured engagement, beginning with a Discovery Assessment and moving through three execution phases, ensuring that each architectural decision was validated against real data and real operational needs before deployment.
The technical architecture begins with ingestion and preprocessing of seven sensor types: temperature, humidity, TVOC, CO₂, PM1, PM2.5, and PM10. Rules-based checks first flag candidate anomalies, then a structured domain-specific prompt — built around HVAC and IAQ framing with a strict JSON output schema — sends the sensor context to Claude Sonnet 4 on Amazon Bedrock. The LLM returns a complete diagnostic package: a natural language root-cause explanation, two to four ranked causes with supporting rationale, specific recommended actions, a confidence score between 0 and 1, and a concise ≤30-word operational summary formatted for direct ticketing system integration.
The most significant architectural innovation came through a mid-project pivot from Phase 1 to Phase 2. In Phase 1, the LLM was used to enhance the outputs of existing rules-based Action Cards, an approach that client stakeholders felt was too close to the original system. In Phase 2, Avahi rearchitected the solution so that the operational rules were injected directly into the LLM prompt, allowing Claude to reason end-to-end from raw sensor context rather than simply elaborating on pre-determined outputs. This rules-as-prompts approach eliminated a middleware layer, ensured AI reasoning stayed consistent with business logic, and produced genuinely differentiated diagnostic outputs that operators could trust and act upon. Average model confidence across evaluated anomalies reached 0.85.
Chain-of-Thought (CoT) prompting was applied throughout to support transparent, traceable reasoning, a critical aspect in regulated environments where operators need to understand why the AI reached a conclusion before acting on it. A human-in-the-loop design ensures that high-risk interventions such as HVAC shutdowns and water system overrides require operator confirmation, keeping accountability with the facility team while reducing the cognitive burden of routine triage. Low-confidence alerts are automatically flagged for human review.
Results are surfaced through a React dashboard hosted on AWS EC2, displaying confidence scores, diagnostic outputs, and a user action feed that allows operators to log responses, creating a feedback loop that will support future model refinement. Sensor data is stored in AWS S3 and Amazon RDS, and all access is governed through AWS IAM with role-based permissions scoped to Attune’s account. The full solution, including prompt libraries and technical documentation, was handed over via GitHub at project completion.
Avahi’s GenAI solution transformed how Attune’s platform interprets and communicates building anomalies, replacing static, generic alert text with dynamically generated, context-aware diagnostics tailored to each event. Facility managers gained the ability to understand root causes and act on specific recommendations without needing deep technical expertise, directly reducing the time and cognitive load required to resolve incidents. Critically, this capability was delivered with the structured reasoning and human-in-the-loop safeguards required for deployment in schools and hospitals — environments where building system failures carry direct health and safety implications.
The measurable outcomes validate the approach:
Let’s explore your high-impact AI opportunities together in a complimentary session
Let’s explore your high-impact AI opportunities together in a complimentary session
Let’s explore your high-impact AI opportunities together in a complimentary half-day session
