Claude Agents on the Shopfloor: OCR, Voice, IoT and Anomaly Detection for Zero-Touch ERP/SAP

The dirty secret of ERP implementations in Indian manufacturing: the software works. The data doesn't get in.

Operators on a running line don't stop to log SAP entries. When they do, the entries are late, incomplete, or wrong. Supervisors backfill at end of shift from memory. Finance gets yesterday's fiction instead of today's reality. Production decisions are made on stale numbers. The ERP becomes a liability instead of an asset.

StratAI identified this pattern across every manufacturing client we've worked with — from textile production at Precot Limited to garment operations at Symphony and Sabrika. The ERP wasn't the problem. The shopfloor data pipeline was.

Claude fixes this at the source.

Three Entry Points — Zero Manual Logging

OCR — Paper to SAP in Seconds

Most shopfloors still run on paper: shift logs, batch sheets, quality checklists, delivery challans. Claude's vision capability reads these documents natively — handwritten or printed, structured or freeform. The OCR pipeline captures the image (via a mounted tablet or phone), Claude extracts the structured data, validates it against expected ranges, and writes it to the correct SAP fields. The operator never touches a keyboard.

Voice Entry — Hands-Free Production Logging

The best interface for a shopfloor is no interface. With Claude's voice entry layer, operators speak production updates in their natural language — Tamil, Hindi, or English. "Machine 4 completed 180 metres, zero defects" becomes a structured SAP production order update in under 3 seconds. Claude validates the entry (is 180m within expected range for this machine and shift?), flags anomalies, and confirms back to the operator via audio.

IoT Integration — Sensor Streams Directly to Claude

Temperature sensors. Motor current monitors. Weigh scales. Vision inspection cameras. All of these produce data that traditionally requires a human to read, interpret, and decide on. Claude connects to these streams directly via MQTT or HTTP and acts as the interpretation layer — not just storing the data, but understanding it in context.

The Anomaly Detection Agent — Act Before It Fails

This is where Claude's intelligence creates genuine competitive advantage.

Traditional monitoring systems alert when something is wrong — a machine trips, a parameter exceeds limit, an alarm fires. By then, you have downtime. You have scrap. You have a customer call to make.

Claude's anomaly detection agent works differently. It maintains a rolling baseline of normal production parameters for every machine, every line, and every product — built from 30, 60, and 90-day historical data stored in Supabase. It compares live sensor data against these baselines continuously.

When a motor's current draw starts trending 6% above its 45-day baseline — before it reaches the alarm threshold — Claude flags it. When a quality metric begins a slow drift that would go unnoticed in day-to-day monitoring — Claude catches it. When the combination of humidity, temperature, and shift timing correlates with a historical defect pattern — Claude predicts it.

Alerts arrive via Telegram to the relevant engineer with full context: which machine, what's anomalous, what the data looks like over the past 4 hours, and what Claude predicts will happen if unaddressed. The engineer acts proactively. The failure never happens.

Why StratAI Builds This Differently

Every shopfloor AI vendor promises connectivity. What they don't do is understand the operational reality of your specific shopfloor before building. StratAI's first step on every manufacturing engagement is a structured operations analysis — mapping every data flow, every manual touchpoint, and every failure pattern. The anomaly detection baselines we build reflect your actual production reality, not a generic industrial template.

That analysis phase is what makes the difference between a system that triggers false alarms every hour and one that surfaces the three signals that actually matter. The technology is Claude. The intelligence is context. The context comes from knowing your operation.