Project MORIA: Automating the Biannual Scrap Rate Review

The Challenge

In manufacturing, scrap rates feed directly into material planning, production costing, and capacity calculations. When rates are stale, every downstream calculation drifts — over-ordering materials, understating production costs, and mispricing capacity.

The standard process corrects these rates twice per year. Without tooling, this correction cycle is manual: analysts extract historical data, estimate correction factors by hand, and enter changes into the production system without a formal review structure or audit trail.

Three specific failure modes drove the MORIA project:

1. No forecasting: Corrections were based on simple averages, ignoring seasonal trends and recent process changes
2. No audit record: There was no structured record of why each rate was set, who approved it, or what the previous value was
3. Succession gaps: When a material number changes (article renamed, replaced, or split), historical production data becomes disconnected — making rate estimation for the new number unreliable

The Solution

MORIA is a standalone Electron + FastAPI desktop application that formalizes the entire correction workflow:

Data Ingestion → Prediction → Human Review → Compliance Output

Predictive Engine

The engine selects the forecasting algorithm based on available historical data:

All predictions include confidence intervals. The UI displays the interval alongside the point estimate so the analyst can assess uncertainty before approving.

Human-in-the-Loop Override Interface

Every proposed rate change is pre-populated with the model’s prediction — not a blank field. The analyst can accept, adjust, or reject each proposal individually. Override reasons are recorded in the audit log.

This design reflects a deliberate principle: the model assists the analyst, but the analyst retains final authority. The system is designed to be trusted precisely because it cannot act without a human decision.

Succession Stitching

Before predictions are generated, production history is matched across material number transitions. This “stitching” step joins the old and new article histories into a continuous series, so the predictive engine sees full data depth even for recently renamed materials.

Compliance Output

After review and approval, the application generates a signed PDF report containing:

• All proposed and approved rates with comparison to previous values
• Override decisions and reasons
• MD5 hash of the dataset used for predictions
• Session metadata (analyst, date, data snapshot version)

The hash enables independent verification that the output was generated from a specific, unmodified dataset.

Technical Stack

• Frontend: Electron (desktop shell), React (UI components)
• Backend: FastAPI (Python, localhost only — no network exposure)
• Prediction: Facebook Prophet, Statsmodels Holt-Winters, custom succession stitching
• Data source: Shared production data cube (BIOME Unified Cube — read-only)
• Audit layer: SQLite for session records, MD5-signed PDF for compliance output

Outcome

MORIA transforms a twice-yearly manual exercise into a structured, auditable workflow that can be completed in a single guided session. The compliance output satisfies audit requirements without additional documentation effort. The predictive engine reduces systematic under- and over-estimation caused by simple average calculations.

The Human-in-the-Loop architecture also serves as a training tool: analysts see model confidence intervals, review historical trends, and develop a data-driven intuition for rate setting over successive review cycles.