Industry
Manufacturing / Industrial
Region
Global — US & Europe
Team Size
Enterprise — 200+ engineers
Volume
1,000s of Work Orders / mo
Platform
Service Cloud + Gemini AI
Delivery
8 weeks · 4 phases
60%
Reduction in manual data entry — field engineers freed from transcribing machine readings by hand
<3s
End-to-end extraction cycle — image uploaded to Work Order updated via async Apex
98%
Field accuracy rate — on structured machine readings in production
3
Machine metric types auto-captured — Power On, Run Time, and Total Time per Work Order
0
Manual re-keying steps remaining — extraction fires automatically on every image upload
8 wks
Audit → Architecture → Build → Go-Live — fully production-deployed

Challenge vs. Solution

Six Operational Problems That Blocked Accurate Machine Data Capture — and Six Fixes That Resolved Them

✕ Before — The Challenges 6 Problems
Manual image transcription at scale — field engineers typed machine readings from uploaded images, a slow error-prone step repeated thousands of times per month.
Variable image quality blocked consistent reading — blurred, angled, and mixed-format images (handwritten labels alongside printed digits) made uniform extraction nearly impossible.
Work Orders sat incomplete — records waited until an engineer manually located and entered Power On, Run Time, and Total Time values from each image.
No structured extraction logic existed — images were stored in SharinPix with no downstream automation connecting them to Salesforce field values.
Apex heap size limits blocked prior attempts — the engineering team had no scalable async architecture; inline processing failed on larger image payloads.
Transcription errors corrupted service histories — manual entry mistakes caused downstream discrepancies in maintenance tracking and machine records.
✓ After — What We Built 6 Fixes
Apex-driven AI extraction pipeline — fires on image upload, calls the Gemini Vision API, and writes parsed values back to the Work Order with no human touch required.
Prompt engineered for variable image quality — handles blurred, rotated, and mixed handwritten/printed inputs, returning structured JSON even on imperfect images.
Configurable field mapping via Custom Metadata — extracted values map to time-based Work Order fields through a maintainable config layer, no code changes needed.
SharinPix-to-Salesforce async trigger architecture — fires asynchronously on image attach events, fully resolving the prior heap size bottleneck.
Queueable Apex for enterprise-scale processing — large image payloads are handled gracefully without governor limit failures across thousands of concurrent Work Orders.
Graceful fallback on invalid AI responses — null or low-confidence extractions flag the record for human review rather than silently writing bad values to the Work Order.

What We Did

Five Phases That Took From Manual Transcription to Automated AI Extraction

Discovery

Mapped the Full Image-to-Field Data Journey

We audited how field engineers were capturing machine readings — from image upload in SharinPix through to manual Work Order entry in Salesforce. We identified 6 points of failure and defined the exact field mapping schema before writing a single line of code.

Process Mapping SharinPix Audit Field Mapping Schema Architecture Blueprint
AI Integration

Connected Gemini Vision API via Salesforce Apex Callouts

We built the REST API callout layer in Apex — constructing the image payload, sending it to Gemini Vision, and parsing the structured JSON response. We engineered the prompt specifically for machine instrument images, handling numeric extraction with high confidence across varied image qualities.

Gemini Vision API Salesforce Apex REST Callout JSON Parsing Prompt Engineering
Architecture

Built Async Processing to Bypass Apex Governor Limits

We replaced the prior synchronous approach — which hit heap size limits on larger image payloads — with a Queueable Apex pattern that processes extraction jobs asynchronously. This gave us a scalable, enterprise-safe architecture capable of handling thousands of concurrent Work Order updates.

Queueable Apex Async Processing Governor Limit Handling SharinPix Trigger
Field Mapping

Wired Extracted Values to Work Order Fields via Custom Metadata

We mapped the AI-extracted time values — Power On Hours, Run Time, Total Time — to the correct Salesforce Work Order fields using a Custom Metadata configuration layer. This means Twopir admin team can update field mappings without a code deployment.

Salesforce Custom Metadata Work Order Objects Dynamic Field Mapping Null Handling
Testing & Go-Live

Validated Across Real Machine Image Samples Before Production

We tested extraction accuracy across a batch of real-world image samples covering blurred gauges, angled shots, handwritten readings, and low-contrast labels. Edge cases with invalid or null AI responses were hardened with graceful fallback logic before the system went live in production.

Apex Test Classes Image Sample Validation Error Handling UAT Production Deploy

Before this, every image meant someone had to stop, read the gauge, type in the numbers, and hope they got it right. Now the system does it the moment the image is uploaded — and the Work Order is already updated by the time the engineer walks away from the machine.

— Twopir Project Lead · Manufacturing Enterprise · 2025

Key Outcomes

What Changed — In Numbers and in Practice

60%

Manual data entry eliminated for time-based machine readings

Field engineers no longer transcribe Power On, Run Time, or Total Time values by hand — extraction fires automatically on image upload.

<3s

End-to-end extraction cycle: image uploaded to Work Order updated

The full AI callout, parsing, and field write cycle completes in under three seconds per image via Queueable Apex async processing.

98%

Extraction accuracy on structured machine readings in production

Gemini Vision achieves 98% field accuracy across standard instrument images — with graceful fallback for edge cases flagged for review.

3 fields

Machine metric types now captured automatically per Work Order

Power On Hours, Run Time, and Total Time are all extracted and mapped in a single AI call — replacing three manual data entry steps.

Zero code

Field mapping changes require no deployment — configurable via Custom Metadata

Admin team can remap AI output to different Work Order fields by updating Custom Metadata records — no Apex change, no deployment window needed.

Day 1

Engineers adopted the system without process change — it just happened

Because extraction triggers automatically on image upload, adoption required zero workflow change from field engineers — the system worked invisibly from day one.

Running a Similar Manufacturing or Field Service Operation?

We offer a free AI automation audit — we'll identify exactly where manual data entry can be eliminated from your Salesforce Work Order flow. Findings delivered in 5 business days, no commitment required.


Technologies Used

The Tools and Techniques Behind This Engagement

Salesforce Service Cloud SharinPix Apex Queueable REST API Callouts JSON Parsing (Apex) Salesforce Work Orders Trigger Framework AI Prompt Engineering Error Handling / Fallback Logic

Turn Your Machine Images into Structured Salesforce Data — Automatically

Twopir has spent 12+ years building Salesforce automation that removes manual effort from industrial and field service operations — delivered for 500+ clients across the US, UK, Australia, UAE, and Canada. We don't just configure; we engineer the solution that fits your workflow.

12+ Years · 500+ Clients · Salesforce Gold Partner · HubSpot Gold Partner · US · UK · Australia · UAE · Canada