The Business Challenge
The business situation the organisation faced before the project began.
1- Data silos across locations , 2- Messy, unstructured source data , 3-Duplicate & inconsistent product records 4-Stockouts & reactive inventory management
Why AI Was the Right Solution
Why AI was the right approach — and what alternatives were considered.
Low-stock detection looks like a simple threshold problem, but it isn't. Static thresholds fail because demand is uneven: the same "10 units left" means different things for a fast-moving SKU in festival season versus a slow mover in a quiet month.
Rules-based alerts would have required manually maintained thresholds per SKU per location — which is precisely the kind of upkeep the client's team was already failing at with their spreadsheets. The problem wasn't a lack of rules; it was that demand patterns shift constantly and no human was going to retune hundreds of thresholds.
An analytics-driven approach that observes actual sales velocity and stock movement patterns adapts by itself. It flags the SKU that will stock out in 4 days at current burn rate, not the SKU that happens to cross an arbitrary line. That reasoning-over-patterns capability is what made AI the right fit — and why the alerts earned user trust when static thresholds wouldn't have.
How the Solution Was Delivered
Discovery, design, development, testing and rollout — the journey, not the tooling.
- 1We delivered a full-stack ERP web application (React.js, Node.js, Express.js, Sequelize ORM, SQLite) that unified inventory, sales and financial ledger workflows across all locations in a single system.
- 2Three components did the heavy lifting:
- 3A Python synchronization engine that ingested 10,000+ raw records from BUSY accounting software and Excel files, achieving 99.8% data ingestion accuracy. It automated the consolidation work the team previously did by hand.
- 4A canonical product data model with SKU deduplication logic that merged duplicate product records across 500+ categories, creating one trusted master record per product.
- 5AI-driven inventory analytics that learned sales velocity and stock movement patterns to detect low-stock conditions automatically, pushing business alerts to users before stock ran out.
- 6The result: a single dashboard where any stakeholder can see true stock positions, sales performance and alerts — replacing days of manual reconciliation with a live view.
Key Technical & Architecture Decisions
Architecture, model selection, workflow and trade-offs.
SQLite over a client-server database. The client's scale and zero-IT-infrastructure reality made SQLite the pragmatic choice — zero administration, reliable transactions, and simple backup. Sequelize ORM keeps the option to migrate to PostgreSQL later without rewriting the data layer.
Canonical schema before migration. We designed the normalized product schema and SKU deduplication rules before writing the sync engine. Tempting as it was to migrate data first and clean later, history shows that never happens — the mess just becomes permanent.
Python for the sync engine, separate from the web stack. Data ingestion, cleaning and reconciliation is a different job from serving a web app. A standalone Python engine let us iterate on parsing BUSY exports and Excel quirks without touching the production API.
Accuracy verification against a gold set. Rather than trusting "it imported without errors," we reconciled migrated totals against manually verified source figures until the 99.8% accuracy bar was met — errors in imports are silent by nature, so verification had to be explicit.
Shadow mode before active alerts. AI alerts went through a silent validation phase where they were logged but not shown to users, and compared against what actually happened. Only after accuracy was proven did alerts go live — this sequencing bought user trust cheaply.
Challenges & How They Were Solved
Obstacles hit along the way and how they were overcome.
Messy source data was the first wall. BUSY exports and years of branch Excel files had inconsistent naming, missing fields, merged cells and contradictory stock figures. The solution was the gold-set approach: manually verify a representative baseline, then tune the ingestion pipeline until system and source reconciled.
Deduplication ambiguity. Some records genuinely couldn't be resolved by rules alone — "Colgate 200g" in one branch and "Colgate paste 200gm" in another might or might not be the same item. Rather than guessing at scale, we flagged ambiguous matches for human review and merged only high-confidence duplicates automatically. The remaining judgment calls were made by people who knew the catalogue.
Alert fatigue risk. Early low-stock thresholds produced too many notifications, and we knew users would start ignoring them. We tuned sensitivity per category and measured alert precision during shadow mode, accepting slightly fewer alerts in exchange for every alert being worth acting on.
The Excel habit. Staff had years of muscle memory around personal spreadsheets. Forcing immediate abandonment would have created resistance and parallel shadow records — so we ran both in parallel during rollout until the system demonstrably matched what their spreadsheets said, at which point the spreadsheets retired themselves.
User Adoption & Change Management
How users responded — training, change management and feedback loops.
Adoption strategy centered on one insight: the team had been burned by software before that showed wrong numbers, so trust had to be earned with their own data, not promised with a demo.
We ran a shadow-mode phase where AI alerts were generated silently and checked against reality. When the first visible alerts proved correct — stock really did run out where the system said it would — trust in the alerts followed quickly.
Training was role-based and short: warehouse staff learned stock entry and alerts; the owner learned dashboards and reorder decisions; nobody was trained on features outside their job. (Adjust to what you actually did.)
The decisive moment was the first monthly close done without the Excel consolidation ritual. When the owner saw the period numbers appear in the dashboard in minutes instead of days, the system stopped being "the new tool" and became the way the business runs.
Feedback loops stayed open throughout: wrong product matches, misclassified SKUs and alert misfires were fixed within days, which kept users reporting issues instead of working around them.
Business Outcomes & Impact
The measurable outcomes from the underlying AI Deployment — with the story behind them.
"ERP with AI inventory intelligence — 99.8% accurate data ingestion, zero duplicate SKUs, 35% fewer stockouts."
Business Outcome Categories
The story behind the numbers
The deployment's headline outcomes were accuracy and availability:
• 99.8% ingestion accuracy across 10,000+ raw records migrated from BUSY accounting exports and Excel — verified against source totals during UAT (confirm how you verified).
• 35% reduction in stockout incidents after AI low-stock alerts went live, measured against the pre-deployment baseline period (insert the baseline period you compared, e.g., "the preceding 3 months").
• Duplicate inventory records eliminated across 500+ product categories, replacing multiple conflicting records per product with a single canonical SKU.
• Manual reconciliation effort removed — consolidation of multi-location stock, sales and ledger data is now continuous and automatic rather than a periodic manual task (insert hours per week/month if you know it).
(Honesty note: only keep figures you can explain if reviewed — the platform's evidence standard means vague claims can hold the case study back.)
Lessons Learned
What surprised the team, what worked well, and what would be done differently.
- Ambiguous duplicates need human judgment — flag them, don't guess
- Shadow-mode validation is the cheapest trust-building investment in AI delivery
- The canonical schema decision was the highest-leverage technical choice of the project
- Alert fatigue is an adoption killer — tune precision before recall
- Accuracy claims must be verifiable against a gold set, or they shouldn't be claimed at all
- Users abandon shadow spreadsheets on their own once the system provably matches their numbers
Future Opportunities
Where this solution could go next.
Demand forecasting. The next evolution is moving from "this will stock out soon" to "order this much, this week" — predictive reorder quantities based on sales velocity, seasonality and lead times.
Automated purchase suggestions. Agentic workflows that draft purchase orders for human approval, using supplier history and current stock positions.
Financial intelligence. The sales and ledger data already in the system supports AI-driven margin analysis, slow-moving stock identification and dead-inventory alerts.
Multi-channel integration. Syncing e-commerce and marketplace orders directly into inventory, extending the same deduplication and analytics pipeline to online channels.
Multi-tenant SaaS. The architecture is close to supporting a hosted version for similar trading businesses — the same sync engine and canonical schema would onboard a new client in days rather than weeks.
Professional Reflections
Guidance for another builder tackling a similar problem.
Design the canonical schema before you migrate anything. Every hour spent on schema and dedup rules upfront saved days of cleanup later. Data migrated into a bad structure becomes permanent mess.
Verify accuracy explicitly — imports fail silently. "No errors thrown" means nothing. Reconcile against a manually verified gold set, and say the number out loud (99.8% here) only after you've earned it.
Ship AI in shadow mode first. Trust is earned in silence. Let the system prove itself against reality before users see its output — one visible mistake costs more trust than ten correct predictions build.
Tune for alert precision, not recall. A system that cries wolf gets ignored; a system that alerts rarely but correctly gets acted on. Fewer, better alerts beat comprehensive noisy ones.
Let parallel runs retire the old system. Don't fight the spreadsheet habit head-on. Run both until the new system visibly wins, and adoption becomes the user's own decision — which is the only kind that sticks.