Case Study — YieldSphere

Production Efficiency Dashboard

Increasing Processing Throughput by 18% Without Capital Investment

$4.5M additional processing capacity through real-time bottleneck intelligence

Industry

Agricultural Processing

Timeline

14 weeks

Team

5 engineers

Tech

IoT + TimescaleDB + React

The Challenge

A large agricultural processing facility operating 3 production lines was running at a claimed 87% capacity utilization — but actual throughput was 23% below rated capacity. Capacity planning relied on annual time-and-motion studies. No real-time visibility into bottlenecks, stoppages, or quality-induced rework cycles existed at the line level.

Our Approach

How We Solved It

Processing Line Instrumentation

Deployed IoT sensors at 48 key process points across 3 lines measuring flow rates, equipment status, quality parameters, and reject rates — establishing the first real-time process data foundation the facility had ever had.

Bottleneck Detection Engine

Built a real-time bottleneck detection algorithm using queuing theory to identify the current constraint limiting each line's throughput, updating every 30 seconds as line conditions change.

Shift Performance Analytics

Developed shift-level performance analytics that decompose the gap between actual and rated throughput into attributable causes: planned downtime, unplanned downtime, speed losses, and quality rework — enabling targeted improvement actions.

Throughput Forecasting

ML model predicts end-of-shift and end-of-day throughput from early-shift process data, giving production managers actionable 4-hour lead time to adjust staffing, schedule maintenance, or expedite raw material deliveries.

Engineering Process

How We Built It

Real-Time Queuing Theory Application

The bottleneck algorithm implements Little's Law and queue-based throughput analysis on the live sensor stream, identifying which station's utilization is constraining system throughput at any moment.

TimescaleDB for Process Time-Series

TimescaleDB continuous aggregates pre-compute shift and daily process summaries without impacting real-time query performance — analytical queries return in under 500ms despite the 48-sensor, 30-second resolution data volume.

Configurable Alert Thresholds

Alert thresholds for each process parameter are configurable by shift supervisors within policy limits — giving operators ownership of their process while maintaining safety boundaries.

Architecture Decisions

Key Technical Choices

Bottleneck as Dynamic, Not Static

Traditional capacity planning treats bottlenecks as fixed by design. Our real-time bottleneck detection revealed that the constraint changes 3-6 times per shift — a finding that invalidated all previous capacity improvement projects.

Causal Attribution Before Correlation Analysis

We built process knowledge into the system (machine → input → output causal model) before adding statistical analysis, avoiding the classic 'correlation without causation' trap in unguided ML on process data.

Supervisor-Facing First, Management-Facing Second

The primary dashboard is optimized for shift supervisors making real-time decisions, not for management reporting. Management views are aggregations of supervisor-level data — not separately designed views.

Results

What We Delivered

18%

Throughput Increase

$4.5M

Additional Annual Capacity Value

Zero

New Capital Investment

99.4%

Process Data Capture Rate

Solution Blueprint

How It All Fits Together

Sensor & IoT Layer

48 process instrumentation points
Edge data aggregation
Real-time event stream

Analytics Layer

Real-time bottleneck detection
OEE decomposition
End-of-shift throughput forecasting

Operations Layer

Shift supervisor dashboard
Maintenance exception alerts
Shift performance reports

Lessons Learned

What We Improved

The Dynamic Bottleneck Discovery Changed Everything

The finding that the bottleneck shifts 3-6 times per shift invalidated 5 years of capital investment planning based on static bottleneck assumptions. The most valuable output of the project was this insight, not the dashboard.

Supervisors Need Actionable, Not Informational

Early dashboard prototypes showed too much data. Supervisors needed 3 things: current bottleneck, severity, and recommended action. Removing 80% of the initial dashboard metrics increased adoption by 4x.

Throughput Improvement Is Organizational, Not Technical

The platform identified 12 bottleneck causes. 8 required no technology — they were scheduling, staffing, and maintenance practice changes. The technology made the problem visible; the operations team solved it.