Sai Computers Ltd Cuts Billing Errors and Revenue Leakage with an AI Meter-Reading Verification Engine on AWS

Company Overview

Sai Computers Ltd (SCL) provides IT and operations services to India’s power distribution sector. It began as an IT consultant for automated batch billing at UPSEB, then expanded into transformer manufacturing, GIS software, handheld spot billing, and AMR/CMRI-based billing and data analysis. Today SCL handles meter reading, billing and energy accounting for discoms across several states. It also runs as a distribution franchisee in Meghalaya, where it is directly accountable for cutting AT&C losses.

Customer Challenge

In SCL’s spot billing operations, meter readers visit consumer premises with handheld devices, key in the meter reading, and capture a photo of the meter as evidence. With very large reading volumes every billing cycle, SCL’s audit team could manually verify only a small fraction of photos. Common problems slipped through: digits mis-keyed by readers, readings estimated without a visit, blurred or unrelated photos, “premises locked” entries used to avoid visits, and signs of meter tampering. These errors led to incorrect bills, consumer complaints and billing disputes, delayed collections, and revenue leakage that contributed to aggregate technical and commercial (AT&C) losses. Manual verification couldn’t scale with volume, and adding auditors was expensive. SCL needed an automated, accurate, cost-effective way to verify every reading before bills were finalised, without disrupting its existing billing workflow or field devices.

Solution Approach

CloudTry designed and built an event-driven, serverless verification engine in SCL’s AWS account in the Asia Pacific (Mumbai) Region, keeping utility and consumer data in India. Handheld devices upload meter photos with reading metadata, including consumer ID, meter number, entered reading, GPS coordinates, and timestamp, to Amazon S3 through Amazon API Gateway using pre-signed URLs. Each upload triggers an AWS Step Functions workflow orchestrated by AWS Lambda.

The workflow first runs image-quality checks for blur, darkness, and whether a meter is visible. Amazon Rekognition text detection and a multimodal model on Amazon Bedrock (Anthropic Claude) then extract the displayed reading, meter serial number, and display type, including digital LCD and electromechanical dial meters, and flag visual signs of tampering such as broken seals, bypass wiring, or covered displays. The extracted reading is compared with the value entered by the reader. Lambda also checks the reading against the consumer’s consumption history stored in Amazon Aurora PostgreSQL, using rules for negative consumption, sudden spikes or drops, and repeated identical readings, and verifies GPS distance from the registered premises.

 

Each reading is scored and classified as verified, mismatch, poor-quality photo, suspected tampering, or abnormal consumption. Verified readings flow straight to SCL’s billing system through an API. Flagged readings go to a review queue in a web dashboard hosted on Amazon S3 and Amazon CloudFront and secured with Amazon Cognito, where auditors confirm or correct them before bills are issued. Amazon QuickSight dashboards show exception rates by area, feeder, and meter reader to support energy accounting and targeted field inspections. Amazon CloudWatch provides monitoring and alerting, AWS KMS encrypts data at rest, and IAM least-privilege roles secure every component. Infrastructure is defined in AWS CDK and deployed through AWS CodePipeline.

Outcomes & Impact

SCL now verifies every meter reading automatically, compared with the small manual sample checked before. Mis-keyed and estimated readings are caught before bills are issued, which reduces billing errors and consumer complaints. Auditors review only the flagged exceptions instead of searching through photos, so the same team covers far more volume. Tampering and abnormal-consumption alerts give field teams a prioritised list for inspections, supporting SCL’s efforts to reduce AT&C losses. Reader-level exception reports have improved field discipline, and the pay-per-use architecture scales smoothly at billing-cycle peaks without idle servers between cycles.

Conclusion

CloudTry compared three options over three years: expanding the manual audit team to verify a larger share of readings, deploying a self-hosted OCR and machine learning stack on dedicated servers, and a managed serverless solution on AWS. The analysis covered staffing, hardware and licensing, implementation effort, infrastructure, model maintenance, and operations. Manual auditing scaled linearly with reading volume and still couldn’t cover all readings. A self-hosted stack needed upfront hardware, sizing for peak billing-cycle loads, and specialist ML operations skills. Billing workloads are highly cyclical, and the AWS solution charges only for images processed, with Lambda, Step Functions, Rekognition, and Bedrock billed per use, so there is no idle cost between cycles. Estimates were built with the AWS Pricing Calculator and validated against actual AWS billing after go-live, and the AWS option delivered the lowest three-year TCO. Revenue recovered from corrected readings and reduced dispute handling was included as a benefit.