Case study / Machine learning

ML Product Mapping System

A multi-model machine learning ensemble (PyTorch Bi-Encoder, Cross-Encoder, XGBoost, Random Forest, & Online Active Learner) that automatically maps distributor descriptions across DBC Group brands (Rucika, Djabesmen, RB Shera, Superex) to internal Master SKUs with 95.4% precision.

PythonPyTorch Bi-EncoderCross-Encoder TransformerXGBoostRandom ForestScikit-Learn (SGD Active Learner)TF-IDF & Fuzzy Matching

Problem

Thousands of unstandardized supplier Purchase Orders and invoices arrived with inconsistent dimension formats (1/2" vs 0.5", AW vs D pipe classes, threaded vs plain fittings). Manual reconciliation across tens of thousands of SKUs took days each month and caused costly fulfillment errors.

Data

The working records and data signals are described in the local project narrative below.

Approach

Engineered a 6-component hybrid hierarchical ensemble combining 384-dimensional Bi-Encoder retrieval (~3ms), Cross-Encoder transformer re-ranking, XGBoost, Random Forest, and a sub-millisecond Online Active Learning feedback loop.

System

0101. Fast Dense Retrieval

PyTorch Bi-Encoder (384-d dense embeddings) retrieves Top 30 candidate Master SKUs in ~3ms

0202. Deep Semantic Re-Ranking

Cross-Encoder Transformer conducts word-level cross-attention for sensitive specs (AW vs D, drat)

0303. Hybrid Meta-Classifier

XGBoost + Random Forest compute probabilities across 8 statistical & lexical feature dimensions

0404. Active Human-in-the-Loop

Online Active Learner (SGD log_loss) updates ensemble weights in <1ms from operator feedback

NOTE

Executive Summary & System Architecture:

- Core Challenge: Ingestion of unstandardized supplier Purchase Orders with highly noisy shorthand (*"RCK PPA AW 1/2 IN"* vs *"PIPA PVC RUCIKA STD AW 0.5 INCH"*) caused severe fulfillment mismatches and 3-day manual audit backlogs.

- Technical Solution: Built a 6-component hierarchical ML ensemble combining 384-dimensional dense vector search (~3ms), Cross-Encoder transformer re-ranking, XGBoost, Random Forest, and an online sub-millisecond active learning loop.

- Quantified Impact: Reached 95.4% precision on auto-approved mappings, automated >80% of PO lines with zero manual intervention, and slashed month-end reconciliation time by 99.1% (from 3 days to <5 minutes) across 4 major industrial brands.

01. Problem Context: Fragmented Distributor Taxonomy

Across DBC Group (Djabesmen Group), thousands of Purchase Orders and invoice lines arrive daily from independent distributor networks for 4 manufacturing brands:

Rucika: PVC pipes (Class AW/D), PPR/HDPE pipes, and specialized plumbing fittings.
Djabesmen: Corrugated fiber-cement roofing sheets and ridge cap accessories.
RB Shera: Fiber cement decorative boards and wood-grain textured planks.
Superex: Rainwater drainage gutter systems and industrial PVC fittings.

Critical Operational Bottlenecks:

Unstructured Input Shorthand: Inconsistent abbreviations (1/2" vs 0.5", AW vs D pressure ratings) caused manual spreadsheet mapping delays.
High-Risk Specification Mismatch: Conflating high-pressure Class AW pipes with thin Class D drainage pipes caused critical logistics dispatch errors.
Manual Overhead: Required 3 full business days at month-end to reconcile inventory cross-references.

02. The 5-Model + TF-IDF Hierarchical Ensemble

To achieve sub-50ms latency with 95%+ precision, the system implements a weighted hierarchical decision pipeline:

Mathematical Model • Econometric FormulationSPECIFICATION

Final Score = 0.20 × M₁ + 0.25 × M₂ + 0.20 × M_3 + 0.15 × M_4 + 0.10 × M_5 + 0.10 × TF-IDF

Component	Model Architecture	Latency	Specialized Operational Role
Model 1 (20%)	PyTorch Bi-Encoder	~3ms	384-d dense embeddings for fast candidate retrieval (Top 30 SKUs).
Model 2 (25%)	Cross-Encoder Re-Ranker	~18ms	Transformer cross-attention evaluating fine-grained specs (AW vs D, threading).
Model 3 (20%)	XGBoost Classifier	~2ms	Gradient-boosted decision trees over 8 lexical and statistical features.
Model 4 (15%)	Random Forest (100 Trees)	~2ms	Bagged meta-ensemble regularizing sparse or rare SKU variants.
Model 5 (10%)	Online Active Learner	<1ms	Incremental `SGDClassifier` absorbing operator feedback in real time.
TF-IDF (10%)	Character N-Gram Vectorizer	<1ms	Sub-word character matching to handle severe distributor typos.

03. Operational Workflow: 3-Tier Confidence Triage

MERMAID
10 LINES

flowchart TD
    A["Distributor PO / Invoice Text"] --> B["Text Preprocessing & Dimension Normalization<br/>(Regex inch-mm, brand aliases, specs)"]
    B --> C["Hierarchical Ensemble Scoring<br/>(Bi-Encoder ➔ Cross-Encoder ➔ XGBoost ➔ RF)"]
    C --> D{"Confidence Triage"}
    D -->|"Score ≥ 85%: Auto-Approve"| E["ERP SAP Master SKU<br/>(>80% PO Volume, 95.4% Precision)"]
    D -->|"Score 60-85%: Review Queue"| F["Human Review Queue<br/>(Top 3 Suggestions)"]
    D -->|"Score < 60%: Anomaly Flag"| G["Master Data Triage<br/>(New Uncataloged SKU)"]
    F --> H["Operator Validates Choice"]
    H --> I["Online Active Learner<br/>(Incremental SGD Update <1ms)"]
    I --> E

Confidence Tier Specifications:

Tier 1: Auto-Approved (Score ≥ 85%): Covers >80% of daily PO volume, directly synchronizing with the SAP ERP system with 95.4% precision.
Tier 2: Human Review Queue (Score 60%–85%): Surfaces the Top 3 recommended candidates. Operator clicks update the online active learner in <1ms.
Tier 3: Anomaly / Uncataloged Flag (Score < 60%): Isolates new unreleased product lines for Master Data team cataloging.

04. Multi-Brand Transformation Matrix

Production transformation examples across DBC Group divisions:

Brand	Raw Distributor Description (Input)	Standardized Master SKU (Output)	Score	Status
Rucika	`RCK PPA AW 1/2 INCH X 4M PUTIH`	`RUCIKA-PVC-AW-050-4M-WHT (Pipa PVC Standard AW 1/2" 4M)`	98.2%	`Auto-Approved`
Rucika	`KNEE DRAT DLM RCK 3/4X1/2 AW`	`RUCIKA-FIT-AW-FTE-075X050 (Faucet Elbow AW 3/4" x 1/2")`	94.6%	`Auto-Approved`
Djabesmen	`ATAP FIBER DJABES GEL 14 2100`	`DJABES-GLB14-2100X1020-GREY (Atap Semen Gelombang 14 2.10M)`	96.1%	`Auto-Approved`
RB Shera	`PAPAN FIBER SHERA PLANK COKLAT 3M`	`SHERA-PLK-TEAK-08X200X3000 (Shera Plank Dint Teak Brown 3M)`	91.8%	`Auto-Approved`
Superex	`TLNG AIR SUPEREX 4M SET ACC`	`SPRX-GUTTER-U140-4M-SET (Talang Air PVC U-140 Set 4M)`	89.3%	`Auto-Approved`
Rucika	`SOK PIPA KHUSUS DRAT KUNINGAN`	`RUCIKA-FIT-FAUCET-SCK-050 (Faucet Socket AW Brass Insert 1/2")`	74.5%	`Reviewed (Top 1)`

05. Measurable Enterprise Business Impact

Operational Metric	Before ML Automation	After ML Deployment	Enterprise ROI
Batch Reconciliation Time	3 Full Business Days	< 5 Minutes	99.1% Turnaround Reduction
Automated PO Volume	0% (100% manual review)	> 80% Zero-Touch	4.2 FTE Labor Hours Saved Daily
Auto-Approved Precision	N/A	95.4% Audited Precision	Zero Schedule AW/D Mismatches
Model Retraining Downtime	Periodic Batch Retraining	< 1ms Online Incremental Update	Zero Operational Interruption

06. Strategic Machine Learning Lessons

Ensemble Diversity Over Single Models: Combining dense embeddings with decision trees and character n-grams overcomes the limitations of any single NLP model.
Calibrated Confidence Tiers: Enforcing an explicit human-in-the-loop review queue for ambiguous items builds organizational trust and safeguards inventory accuracy.
Sub-Millisecond Active Learning: Continuous incremental learning from daily operator corrections prevents repeating identical manual fixes.

Impact

Achieved 95.4% precision on auto-approved mappings, automated >80% of PO lines without manual touch, and slashed batch processing turnaround from 3 business days to under 5 minutes across 4 manufacturing brands.

Lessons

Multi-model ensembles provide significantly greater robustness against extreme technical phrasing variations than any single NLP model.
Strict confidence tiering (Auto-Approve vs Human Review Queue) is critical for enterprise operational trust and zero-downtime adoption.
Online incremental learning ensures continuous model improvement without requiring heavy, disruptive batch retraining pipelines.