Recall on pole anomalies, well above the 80% acceptance bar
More accurate tilt forecasts than the seasonal baseline
Of risk forecast for every pole, refreshed daily and during storm watches
N-Smart Inc.
Reston, Virginia
Infrastructure monitoring for utilities and municipalities
Amazon SageMaker, Amazon Bedrock (Anthropic Claude Haiku 4.5), AWS Lambda, Amazon API Gateway, Amazon DynamoDB, Amazon S3, Amazon CloudFront, Amazon Simple Notification Service, Amazon EventBridge, Amazon Elastic Container Registry
N-Smart Inc. provides infrastructure monitoring solutions for utilities and municipalities. Its pole monitoring system integrates sensor telemetry with environmental and weather data to track tilt, vibration, and surrounding conditions across a fleet of utility poles. To prevent structural failures, outage events, and wildfire ignition, N-Smart engaged Avahi to build an AWS native analytics and GenAI solution that scores each pole for structural and fire risk in real time, projects that risk forward across the upcoming week, and explains the results in plain English for the operations team. The solution runs entirely inside N-Smart’s AWS account and now feeds their operations workflow with continuously refreshed risk scores, forward-looking forecasts, and prioritized recommendations.
N-Smart Inc., headquartered in Reston, Virginia, delivers connected infrastructure monitoring for utility operators and municipal customers. The company’s platform aggregates data from purpose-built sensors deployed on physical assets in the field, combines that telemetry with environmental context, and translates the joint signal into operational intelligence that helps utilities prevent equipment failure, service outages, and public safety incidents. N-Smart’s existing platform includes a real-time telemetry pipeline that ingests sensor readings from purpose-built devices attached to poles and stores the data in a partitioned data lake on AWS.
Utility operators face two operational risks that are both growing in frequency and severity: wildfire ignition originating at aging or damaged equipment, and storm-driven outages that can affect thousands of customers at once. Both are difficult to prevent because they depend on the joint behavior of many variables at once: the physical condition of the pole, the intensity of local weather, the moisture in the soil, the density of surrounding vegetation, and the accumulated impact history of the asset. Reading each variable in isolation misses the interactions that actually cause failures.
N-Smart had the raw data. Their existing telemetry pipeline captures sensor readings on every uplink, and the readings land in a partitioned data lake ready for analysis. What N-Smart needed was the analytics and reasoning layer that turns those readings into risk scores an operations team can act on: a score for each pole today, a forward-looking projection across the upcoming week, and a plain-English explanation the on-call operator can read at three in the morning to decide whether to dispatch a field crew.
Without this layer, N-Smart’s operations teams would continue to work reactively, responding to failures after they occurred rather than intervening before them. For a utility running thousands of poles across a service territory, that gap translates directly into avoidable service interruptions, higher insurance exposure from wildfire liability, and slower response to storm events that could otherwise be pre-staged. N-Smart wanted a solution that would let their customers move from reactive maintenance to predictive risk management, and they wanted it delivered as production-grade AWS infrastructure their internal team could own and extend.
N-Smart’s existing telemetry pipeline was already deployed on AWS, with real-time ingest, a partitioned data lake, and identity and access management in place. Building the analytics and reasoning layer on the same platform kept the solution close to the data, avoided cross-cloud egress costs, and let N-Smart apply the same operational tooling they already use for compliance, monitoring, and cost governance.
AWS also offered the specific managed services the solution required, each of them production ready and covered by AWS enterprise support: Amazon SageMaker for anomaly detection and time-series forecasting, Amazon Bedrock for large language model access, and a mature serverless stack for the API, alerting, and dashboard delivery. Anchoring the new work on the AWS platform meant N-Smart could scale from the initial fleet to their broader customer base without a platform migration.
N-Smart selected Avahi based on the successful delivery of an earlier engagement that built their existing telemetry pipeline on AWS. Avahi’s track record as a premier tier AWS partner and its concentration of experience across GenAI, SageMaker, and production data engineering gave N-Smart confidence that the same team could take the platform from raw data ingest to operational risk intelligence in a single continuous engagement.
Avahi’s approach also matched what N-Smart needed as an infrastructure monitoring provider. Every score in the final solution traces back to a specific contractual line item or to a section of the customer’s own requirements specification, so the operations team can explain to their own customers exactly why an alert fired. Avahi’s engineering team paired closely with N-Smart on a weekly cadence through the engagement, and each iteration of the scoring formulas and thresholds landed against operational rules the two teams agreed to together.
Avahi and N-Smart built an AWS native analytics and explainability pipeline that runs entirely inside N-Smart’s AWS account. The pipeline is organized around three layers: a real-time scoring engine that runs on every sensor uplink, a daily forecast engine that projects the next seven days of risk per pole, and a read API that serves an operator dashboard.
The real-time scoring layer runs on AWS Lambda. On every uplink from a pole, the enrichment function reads the sensor reading from the telemetry data lake in Amazon S3, joins it with external environmental data (weather from the National Weather Service and Open-Meteo, soil moisture from NASA SMAP, and vegetation density from NASA MODIS), computes a set of derived signals, and writes an enriched row that carries every downstream score. Two rule-based detectors fire against the enriched row: a tilt alert with cause attribution (wind, impact, ground failure, or unknown) and a motion event anomaly detector that compares each pole against its own thirty-day baseline. In parallel, an Amazon SageMaker Random Cut Forest model looks at the joint behavior of six sensor features and flags multivariate anomalies that no single rule would catch. Three composite scores ride on the same row: an Ignition Risk Index built from six normalized inputs using the National Fire Danger Rating System framework, a Pole Health Score that blends structural stress and equipment load, and a Regional Fire Vulnerability score that rolls the per-pole risk up to a regional daily view.
The forecast layer runs once per day per pole through a scheduled Lambda triggered by Amazon EventBridge. It reads the trailing thirty days of enriched history and scores three forecast metrics (tilt angle, acceleration magnitude, and temperature) through an Amazon SageMaker DeepAR model. DeepAR was chosen because it learns a single embedding shared across every pole in the fleet, which means one trained model serves every device and new poles join the embedding automatically through the weekly retrain. The Lambda projects each metric out one hundred sixty-eight hours as p10, p50, and p90 confidence bands, derives operational risk numbers from the bands (the probability of tilt crossing the sustained-lean threshold at twenty-four, forty-eight, and seventy-two hours; an Ignition Risk Index projection across the week; and a Storm Outage composite probability), and caches the result in Amazon DynamoDB for the dashboard to read. When any operational risk crosses a configurable threshold, the Lambda publishes an alert through Amazon Simple Notification Service so the on-call operator receives the notification by email without needing to watch the dashboard.
The explanation layer runs on Amazon Bedrock. When the operator opens a pole page on the dashboard, the API calls Bedrock with the pole’s latest reading, cached forecast bands, current scores, and regional context. The model returns a short paragraph in plain English plus a prioritized list of imperative actions the on-call operator can take in the next few hours. The explanation layer supports five distinct query patterns tuned to the two core use cases, wildfire ignition and storm outage, and enforces a structured response contract that the dashboard validates before rendering.
The delivered read API is exposed through Amazon API Gateway and supports six JSON endpoints: the pole list, the current reading and cached forecast for a pole, trailing history, the forecast bands, the natural-language summary from Bedrock, and the daily regional roll-up. A Vue single page application built with Vite is served through Amazon CloudFront in front of a private S3 origin and gives N-Smart’s operations team a single place to review alerts, forecasts, reasoning, and Bedrock summaries per pole. The dashboard and the JSON endpoints are the same interface the operations team uses today, and both are equally callable from an operations tool for integration into N-Smart’s existing systems. A continuous integration and continuous deployment pipeline built through Amazon Elastic Container Registry handles versioned rollouts of the Lambda functions.
Real-time scoring engine on AWS Lambda that enriches every sensor uplink with weather, soil moisture, and vegetation context and writes a fully scored row
Two rule-based detectors (tilt alert with cause attribution and motion event anomaly) with configurable severity bands and structured reason records
Amazon SageMaker Random Cut Forest model for multivariate anomaly detection across six sensor features, retrained weekly
Amazon SageMaker DeepAR forecasting model producing p10, p50, and p90 confidence bands across a seven-day horizon on tilt, acceleration magnitude, and temperature
Composite risk scores: Ignition Risk Index, Pole Health Score, and Regional Fire Vulnerability, each with documented formulas, weights, and validation plans
Amazon Bedrock explainability layer with five query patterns generating plain-English summaries and prioritized operator recommendations
Read API on Amazon API Gateway with six JSON endpoints for poles, current readings, history, forecasts, natural-language summaries, and regional roll-ups
Vue operator dashboard on Amazon CloudFront with five per-pole views (status, forecast, reasoning, AI summary, raw data) plus a regional heat-map view
Alert distribution through Amazon Simple Notification Service across three topics covering real-time alerts, forecasted events, and regional recompute triggers
Continuous integration and deployment pipeline through Amazon Elastic Container Registry for versioned Lambda rollouts
Structured reason records tied to five named failure signatures (imminent collapse, progressive lean, impact damage, sustained motion stress, and inverted or fallen pole)
Full technical handover documentation covering architecture, models, configuration, and API reference
The delivered solution moves N-Smart’s operations from reactive to predictive. Every pole in the deployed fleet is now scored on every sensor uplink for structural and fire risk, a daily forecast projects each pole’s risk seven days ahead, and Bedrock generates plain-English explanations with prioritized actions on demand. Every score in the family traces back to a specific contractual requirement or a clause in N-Smart’s operational specification, which means the operations team can explain the reasoning behind every alert to their own customers.
Random Cut Forest anomaly detection achieved a recall of 0.9758 on a clean validation cohort of 4,975 records, clearing the acceptance threshold of 0.80 by more than nineteen percentage points
DeepAR tilt forecasting reached a Mean Absolute Scaled Error between 0.51 and 0.60, beating the seasonal baseline by 40% to 50%
Forecast hit-rate at the 24, 48, and 72 hour horizons cleared the acceptance threshold of 80% with margin on both tilt and acceleration magnitude
API response latency runs well inside the 10 second acceptance criterion across all endpoints
One hundred sixty-eight hour (7 day) forecast horizon on every pole, refreshed daily and within 15 minutes of any active storm watch
Solution running end to end inside N-Smart’s AWS account, with a versioned continuous deployment pipeline for ongoing evolution
N-Smart Inc.
Reston, Virginia
Infrastructure monitoring for utilities and municipalities
Amazon SageMaker, Amazon Bedrock (Anthropic Claude Haiku 4.5), AWS Lambda, Amazon API Gateway, Amazon DynamoDB, Amazon S3, Amazon CloudFront, Amazon Simple Notification Service, Amazon EventBridge, Amazon Elastic Container Registry
N-Smart Inc. provides infrastructure monitoring solutions for utilities and municipalities. Its pole monitoring system integrates sensor telemetry with environmental and weather data to track tilt, vibration, and surrounding conditions across a fleet of utility poles. To prevent structural failures, outage events, and wildfire ignition, N-Smart engaged Avahi to build an AWS native analytics and GenAI solution that scores each pole for structural and fire risk in real time, projects that risk forward across the upcoming week, and explains the results in plain English for the operations team. The solution runs entirely inside N-Smart’s AWS account and now feeds their operations workflow with continuously refreshed risk scores, forward-looking forecasts, and prioritized recommendations.
N-Smart Inc., headquartered in Reston, Virginia, delivers connected infrastructure monitoring for utility operators and municipal customers. The company’s platform aggregates data from purpose-built sensors deployed on physical assets in the field, combines that telemetry with environmental context, and translates the joint signal into operational intelligence that helps utilities prevent equipment failure, service outages, and public safety incidents. N-Smart’s existing platform includes a real-time telemetry pipeline that ingests sensor readings from purpose-built devices attached to poles and stores the data in a partitioned data lake on AWS.
Utility operators face two operational risks that are both growing in frequency and severity: wildfire ignition originating at aging or damaged equipment, and storm-driven outages that can affect thousands of customers at once. Both are difficult to prevent because they depend on the joint behavior of many variables at once: the physical condition of the pole, the intensity of local weather, the moisture in the soil, the density of surrounding vegetation, and the accumulated impact history of the asset. Reading each variable in isolation misses the interactions that actually cause failures.
N-Smart had the raw data. Their existing telemetry pipeline captures sensor readings on every uplink, and the readings land in a partitioned data lake ready for analysis. What N-Smart needed was the analytics and reasoning layer that turns those readings into risk scores an operations team can act on: a score for each pole today, a forward-looking projection across the upcoming week, and a plain-English explanation the on-call operator can read at three in the morning to decide whether to dispatch a field crew.
Without this layer, N-Smart’s operations teams would continue to work reactively, responding to failures after they occurred rather than intervening before them. For a utility running thousands of poles across a service territory, that gap translates directly into avoidable service interruptions, higher insurance exposure from wildfire liability, and slower response to storm events that could otherwise be pre-staged. N-Smart wanted a solution that would let their customers move from reactive maintenance to predictive risk management, and they wanted it delivered as production-grade AWS infrastructure their internal team could own and extend.
N-Smart’s existing telemetry pipeline was already deployed on AWS, with real-time ingest, a partitioned data lake, and identity and access management in place. Building the analytics and reasoning layer on the same platform kept the solution close to the data, avoided cross-cloud egress costs, and let N-Smart apply the same operational tooling they already use for compliance, monitoring, and cost governance.
AWS also offered the specific managed services the solution required, each of them production ready and covered by AWS enterprise support: Amazon SageMaker for anomaly detection and time-series forecasting, Amazon Bedrock for large language model access, and a mature serverless stack for the API, alerting, and dashboard delivery. Anchoring the new work on the AWS platform meant N-Smart could scale from the initial fleet to their broader customer base without a platform migration.
N-Smart selected Avahi based on the successful delivery of an earlier engagement that built their existing telemetry pipeline on AWS. Avahi’s track record as a premier tier AWS partner and its concentration of experience across GenAI, SageMaker, and production data engineering gave N-Smart confidence that the same team could take the platform from raw data ingest to operational risk intelligence in a single continuous engagement.
Avahi’s approach also matched what N-Smart needed as an infrastructure monitoring provider. Every score in the final solution traces back to a specific contractual line item or to a section of the customer’s own requirements specification, so the operations team can explain to their own customers exactly why an alert fired. Avahi’s engineering team paired closely with N-Smart on a weekly cadence through the engagement, and each iteration of the scoring formulas and thresholds landed against operational rules the two teams agreed to together.
Avahi and N-Smart built an AWS native analytics and explainability pipeline that runs entirely inside N-Smart’s AWS account. The pipeline is organized around three layers: a real-time scoring engine that runs on every sensor uplink, a daily forecast engine that projects the next seven days of risk per pole, and a read API that serves an operator dashboard.
The real-time scoring layer runs on AWS Lambda. On every uplink from a pole, the enrichment function reads the sensor reading from the telemetry data lake in Amazon S3, joins it with external environmental data (weather from the National Weather Service and Open-Meteo, soil moisture from NASA SMAP, and vegetation density from NASA MODIS), computes a set of derived signals, and writes an enriched row that carries every downstream score. Two rule-based detectors fire against the enriched row: a tilt alert with cause attribution (wind, impact, ground failure, or unknown) and a motion event anomaly detector that compares each pole against its own thirty-day baseline. In parallel, an Amazon SageMaker Random Cut Forest model looks at the joint behavior of six sensor features and flags multivariate anomalies that no single rule would catch. Three composite scores ride on the same row: an Ignition Risk Index built from six normalized inputs using the National Fire Danger Rating System framework, a Pole Health Score that blends structural stress and equipment load, and a Regional Fire Vulnerability score that rolls the per-pole risk up to a regional daily view.
The forecast layer runs once per day per pole through a scheduled Lambda triggered by Amazon EventBridge. It reads the trailing thirty days of enriched history and scores three forecast metrics (tilt angle, acceleration magnitude, and temperature) through an Amazon SageMaker DeepAR model. DeepAR was chosen because it learns a single embedding shared across every pole in the fleet, which means one trained model serves every device and new poles join the embedding automatically through the weekly retrain. The Lambda projects each metric out one hundred sixty-eight hours as p10, p50, and p90 confidence bands, derives operational risk numbers from the bands (the probability of tilt crossing the sustained-lean threshold at twenty-four, forty-eight, and seventy-two hours; an Ignition Risk Index projection across the week; and a Storm Outage composite probability), and caches the result in Amazon DynamoDB for the dashboard to read. When any operational risk crosses a configurable threshold, the Lambda publishes an alert through Amazon Simple Notification Service so the on-call operator receives the notification by email without needing to watch the dashboard.
The explanation layer runs on Amazon Bedrock. When the operator opens a pole page on the dashboard, the API calls Bedrock with the pole’s latest reading, cached forecast bands, current scores, and regional context. The model returns a short paragraph in plain English plus a prioritized list of imperative actions the on-call operator can take in the next few hours. The explanation layer supports five distinct query patterns tuned to the two core use cases, wildfire ignition and storm outage, and enforces a structured response contract that the dashboard validates before rendering.
The delivered read API is exposed through Amazon API Gateway and supports six JSON endpoints: the pole list, the current reading and cached forecast for a pole, trailing history, the forecast bands, the natural-language summary from Bedrock, and the daily regional roll-up. A Vue single page application built with Vite is served through Amazon CloudFront in front of a private S3 origin and gives N-Smart’s operations team a single place to review alerts, forecasts, reasoning, and Bedrock summaries per pole. The dashboard and the JSON endpoints are the same interface the operations team uses today, and both are equally callable from an operations tool for integration into N-Smart’s existing systems. A continuous integration and continuous deployment pipeline built through Amazon Elastic Container Registry handles versioned rollouts of the Lambda functions.
Real-time scoring engine on AWS Lambda that enriches every sensor uplink with weather, soil moisture, and vegetation context and writes a fully scored row
Two rule-based detectors (tilt alert with cause attribution and motion event anomaly) with configurable severity bands and structured reason records
Amazon SageMaker Random Cut Forest model for multivariate anomaly detection across six sensor features, retrained weekly
Amazon SageMaker DeepAR forecasting model producing p10, p50, and p90 confidence bands across a seven-day horizon on tilt, acceleration magnitude, and temperature
Composite risk scores: Ignition Risk Index, Pole Health Score, and Regional Fire Vulnerability, each with documented formulas, weights, and validation plans
Amazon Bedrock explainability layer with five query patterns generating plain-English summaries and prioritized operator recommendations
Read API on Amazon API Gateway with six JSON endpoints for poles, current readings, history, forecasts, natural-language summaries, and regional roll-ups
Vue operator dashboard on Amazon CloudFront with five per-pole views (status, forecast, reasoning, AI summary, raw data) plus a regional heat-map view
Alert distribution through Amazon Simple Notification Service across three topics covering real-time alerts, forecasted events, and regional recompute triggers
Continuous integration and deployment pipeline through Amazon Elastic Container Registry for versioned Lambda rollouts
Structured reason records tied to five named failure signatures (imminent collapse, progressive lean, impact damage, sustained motion stress, and inverted or fallen pole)
Full technical handover documentation covering architecture, models, configuration, and API reference
The delivered solution moves N-Smart’s operations from reactive to predictive. Every pole in the deployed fleet is now scored on every sensor uplink for structural and fire risk, a daily forecast projects each pole’s risk seven days ahead, and Bedrock generates plain-English explanations with prioritized actions on demand. Every score in the family traces back to a specific contractual requirement or a clause in N-Smart’s operational specification, which means the operations team can explain the reasoning behind every alert to their own customers.
Random Cut Forest anomaly detection achieved a recall of 0.9758 on a clean validation cohort of 4,975 records, clearing the acceptance threshold of 0.80 by more than nineteen percentage points
DeepAR tilt forecasting reached a Mean Absolute Scaled Error between 0.51 and 0.60, beating the seasonal baseline by 40% to 50%
Forecast hit-rate at the 24, 48, and 72 hour horizons cleared the acceptance threshold of 80% with margin on both tilt and acceleration magnitude
API response latency runs well inside the 10 second acceptance criterion across all endpoints
One hundred sixty-eight hour (7 day) forecast horizon on every pole, refreshed daily and within 15 minutes of any active storm watch
Solution running end to end inside N-Smart’s AWS account, with a versioned continuous deployment pipeline for ongoing evolution
Let’s explore your high-impact AI opportunities together in a complimentary session
