01 · Case study · Data normalization & quality
Turning 2.3 billion dirty data cells into a single source of truth.
How TitanStack AI’s automated normalization pipeline resolved 96% of cross-system discrepancies for Supply Logic, an enterprise supply chain management company — and surfaced a six-figure reconciliation gap along the way.
Last reviewed:
96.1%
Auto-resolved
18.5M
Discrepancies found
701M
Cells cleaned
$115K
Variance surfaced
The client
An enterprise supply chain management company running a core ERP system spread across 12 interconnected source tables — customer, vendor, item, order, invoice, and inventory masters totaling more than 34 million rows of operational data.
The challenge
Years of accumulated data had left the ERP riddled with cross-table inconsistencies. The same customers and vendors were recorded differently from one system to the next, item descriptions were corrupted by stray characters and formatting noise, date fields held malformed and sentinel values (some reading 1800 or even 2221), and pricing figures failed to reconcile against their source calculations. These problems quietly eroded reporting accuracy and day-to-day operational decisions — and no one had quantified the true scale of the damage.
The approach
TitanStack AI built an automated normalization pipeline and ran 27 comparison pairs across the 12 tables, processing every shared data element through three sequential stages:
- Key-based matching joined records across tables on shared keys, flagging anything present in one system but missing from the other.
- Deterministic cleaning removed special characters, noise markers, and whitespace irregularities, then tested for exact equality.
- Fuzzy matching scored remaining near-matches for similarity, auto-collapsing high-confidence pairs and routing only true ambiguities to human review.
Three complementary workstreams ran alongside the pipeline: a full-table regex cleanup touching every cell in the dataset, an independent date-quality pass, and an expenditure recalculation that re-derived invoice line totals from first principles.
The results
- 17.8M records resolved automatically — 96.1% of every discrepancy identified.
- 720K records flagged for manual review (just 3.9%), with vendor codes alone driving 94% of them.
- 701.4M cells standardized across the dataset — nearly a third of all 2.28 billion cells scanned.
- 245,556 dates corrected, including 152K sentinel values hiding in a single item field.
- $115,576.88 in pricing variance surfaced across 617 invoice lines spanning FY2023–FY2026.
The bottom line
With 96% of discrepancies resolved automatically, the client’s team was left with a focused, manageable review queue instead of an unbounded cleanup. And a single recalculation query exposed over $115K in pricing variance — likely a fraction of a far larger reconciliation gap. TitanStack AI recommends extending remediation database-wide and installing upstream controls at the point of ingestion to stop these issues at the source.
What impressed me most about working with TitanStack AI wasn’t just the technology, it was how they operated as an extension of our team… They didn’t just clean our data; they built a pipeline that understood the logic of our business across 12 interconnected systems, then handed us a clear, prioritized path forward instead of a black box. Finding over $115,000 in pricing variance almost as a byproduct tells you everything about the depth of the work. This is what we look for in an AI strategic partner: judgment, not just automation.
What’s hiding in your data?
Book a strategy callWe'll compare notes and see whether an assessment or a build makes sense for you.