WorkCenter Central, a business management platform built from scratch to replace the distributor's previous system: end-to-end workflow automation, QuickBooks integration, and real-time reporting, cutting manual errors 80% and speeding order processing 50% over a partnership now past seven years. An AI SKU price finder was added on top, indexing thousands of historical vendor offer files without changing the core platform.
Lorecs, a distributor of brand-name products to wholesalers and retailers across the United States, had a business management platform that was holding the business back. It depended on manual processes that caused errors, its integrations with accounting, scheduling, and alerting were fragmented, and it couldn't scale with growing distribution volume.
In 2018 our founding engineers took over the platform and, rather than modernizing what was there, built WorkCenter Central from scratch to replace it: a business management platform designed around Lorecs' actual workflows, with an API-first architecture, end-to-end automation from order to invoice, and QuickBooks, scheduler, and alert integrations built in from day one. Seven years on, the same engineers still run it under Rorix, proactively suggesting improvements and delivering consistently on time.
The latest phase put AI to work on data WorkCenter Central had been quietly accumulating for years. Thousands of vendor offer files, Excel and CSV price lists from many different suppliers, sat on disk with no way to compare prices across them short of opening each one by hand. Rorix built a standalone AI SKU Price Finder: a NestJS and MongoDB service that backfilled every historical file, keeps syncing new ones, reconciles the same product across differently formatted vendor lists, and answers questions like 'best price for this SKU' or 'which vendor is cheapest for this brand' in plain English. It shipped without a single change to WorkCenter Central's production codebase.
The platform Lorecs was running on left order entry, vendor follow-ups, and inventory updates to manual work, creating costly bottlenecks and compounding errors that grew with every increase in distribution volume.
QuickBooks, scheduling, and alerting each ran outside the platform, with no shared record of orders or inventory and no way to move data between them without re-keying it.
The existing system could have been patched module by module, but its architecture could not carry the automation, integrations, or scale the business needed, and packaged tools missed the edge cases of wholesale distribution: vendor-specific pricing, quote comparison, product variants, and out-of-stock fulfillment. The call was to build from scratch.
Whatever was built had to hold up as volume, regions, and currencies grew, and keep running through every release and upgrade without interrupting live operations.
Years of vendor offers had been uploaded to WorkCenter Central as Excel and CSV files and tracked in its database, but never parsed. Answering "who sells this SKU cheapest" meant opening files one at a time, so cross-vendor comparison rarely happened at all.
Every supplier authored its own layout: SKU columns labelled item code, style number, UPC, or EAN; prices labelled cost or wholesale; decorative title rows above the real headers; multi-sheet workbooks; and the occasional corrupt, password-protected, or mislabelled file. The same product was spelled differently in nearly every list.
The AI service had to find and fetch files from WorkCenter Central's production database and file storage without new endpoints, schema changes, or write access, and without adding load that could destabilize the live platform. None of the file tables even carried a timestamp to detect new uploads.
A wrong price is worse than no answer. The assistant had to refuse anything outside vendor pricing, say so when no data existed rather than invent a number, and stay immune to prompt injection from the vendor-authored file content it was built on.
Took over the existing platform and mapped how orders, vendors, inventory, and accounting actually flowed through and around it before writing code, then designed WorkCenter Central's data model and module boundaries around those workflows rather than the old system's structure.
Designed and built a scalable RESTful API layer connecting all business systems, establishing a single source of truth and enabling real-time data flow across the entire platform.
Automated order processing, inventory synchronization, and vendor management workflows end-to-end, eliminating the manual touchpoints responsible for errors and delays.
Connected QuickBooks, automated schedulers, and a real-time alert system directly to the core platform from the first release, creating a unified, intelligent operational ecosystem.
Built real-time reporting dashboards that surface actionable insights on inventory levels, order velocity, and vendor performance, enabling confident, data-driven decisions.
Implemented a blue-green deployment strategy and rigorous staging environment so every release, including the platform's move from .NET Core 2.1 to .NET 9, shipped without interrupting live operations.
Provisioned a least-privilege SQL Server login scoped to exactly the offer, company, brand, and file tables, plus a read-only mount of the file storage. Because the file tables carry no timestamps, ingestion tracks progress with an ID-based cursor per source table: no writes, no API calls, and no changes to WorkCenter Central's .NET codebase.
Ran the historical backfill through a BullMQ and Redis queue in batches, with per-file retries and rate limiting so file-storage I/O never spikes. Cursors advance only after a batch commits, so an interrupted run resumes without skipping files, and a scheduled sync picks up new WorkCenter Central uploads automatically.
Parsed Excel and CSV with SheetJS and csv-parse, detecting the real header row instead of assuming row one. A synonym dictionary maps common column variants, an LLM-assisted fallback handles ambiguous headers, and anything still uncertain is marked for review rather than guessed. Corrupt or unsupported files are skipped with the reason recorded, and every file gets an audit record so the team can see exactly how many made it in.
Resolved every parsed row to one canonical SKU by identifier first (UPC or EAN), then exact normalized name, then fuzzy match, with aliases kept for future matches. Vendors are seeded from the platform's own company records rather than spreadsheet headers, so cross-vendor comparisons compare the right entities.
Every question passes a domain guardrail enforced in code before anything else runs, then a structured intent and entity extraction step that maps to a closed set of parameterized query templates. The model never writes database queries, and the final answer is generated only from the normalized result rows, never from raw file content. An empty result produces "no data" instead of a number.
Shipped as its own NestJS application with JWT authentication, its own chat UI, ops endpoints for ingestion status and per-file audit detail, and the LLM behind a provider interface so the model vendor can change without touching the pipeline.
Rorix Technologies took over our business management platform, refined inventory accuracy, automated workflows, and integrated external tools. They deliver on time, adapt to our needs, and actively suggest improvements. We have seen a huge reduction in manual errors, faster order processing, and increased operational capacity.
Let's build the system your operation actually needs, integrate the tools around it, and put AI to work on the data you already have.
Everything on this topic that lives elsewhere on the site.