Demand Sensing System
A nightly demand-sensing layer pulls order, distributor, weather, and macro signals into a single forecast that rebuilds production schedules across six facilities before the first shift starts — closing a planning gap that used to run more than a week.
The challenge
A regional industrial components manufacturer ran six facilities off a weekly S&OP cycle that couldn't react to what was happening in the market:
- Order book, distributor inventory, and weather data lived in three separate systems, refreshed on different schedules
- Schedule changes required a planner to manually reconcile all three before a revision was even proposed
- By the time a demand shift was confirmed and actioned, it was often 10-14 days old
- Weather-driven demand swings (regional cold snaps, storm-linked stockups) were caught after the fact, not ahead of it
- Each facility planned independently, so a shortage at one site rarely triggered a rebalance at another
Average time from signal to actioned schedule change: 68 hours across the six facilities — and that clock didn't start until a planner noticed the signal in the first place.
How it works
A nightly cycle instead of a weekly one
Rather than replace the ERP or the planners, the system was built to feed them a better starting point every morning:
- 01
Order book, distributor POS/inventory feeds, weather forecast APIs, and macro indicators (fuel, commodity indices) were piped into a shared signal store
- 02
A gradient-boosted forecasting model, retrained weekly, scores demand shifts at the SKU-facility level
- 03
Every night, the model re-runs against the day's fresh signals and proposes an updated production schedule for all six facilities
- 04
Each proposal is ranked by expected impact, so planners see the changes worth acting on first, not a wall of diffs
- 05
Planners approve or override each morning inside the existing scheduling tool — no new system to log into
- 06
Approved changes sync back to the ERP automatically, closing the loop before the first shift starts
What we built
Key capabilities
One nightly source of truth
Order, distributor, weather, and macro signals are reconciled into a single schedule proposal instead of three disconnected feeds.
Facility-aware rebalancing
A shortage signal at one site now automatically surfaces a rebalancing option at the other five.
Ranked, not raw, changes
Planners see schedule changes ordered by expected impact, so the highest-value calls get made first.
No new interface
Recommendations land inside the scheduling tool planners already use, so adoption didn't require a change-management push.
Before vs after
How the planning cycle changed
- Planning cadence
- Weekly → Nightly
- Signal-to-schedule time
- 68 hrs → 43 hrs
- Facility coordination
- Independent → Cross-facility rebalancing
- Weather response
- Reactive → Forecast-driven
Business impact
What it changed
37% less decision lag
(68 − 43) ÷ 68 hours — signal to actioned schedule change, averaged across all six facilities.
Six facilities, one cycle
Every site now plans against the same nightly forecast instead of six separate weekly ones.
20-week delivery
Live in production in 20 weeks, integrated into the scheduling tool planners already used.
Technology stack
“Decision lag isn't fixed by better dashboards — it's fixed by shortening the loop between a signal arriving and a schedule reflecting it. Six facilities now run on the same clock.”
Keep reading