Halcrow provided a cross-functional team embedded directly in FMCG Analytics's product development:

Team structure:

• 1 platform architect (SaaS infrastructure, multi-tenancy, security, scalability)

• 2 data engineers (ETL pipelines, warehouse architecture, data quality)

• 1 ML engineer (operationalize RWA's analytics models, production deployment)

• 1 frontend engineer (self-service UI, data visualization, dashboard building)

• 1 product lead (RWA/Halcrow interface, requirements → roadmap)

Critical embedding elements:

• Weekly product alignment: FMCG Analytics founders, Halcrow team, coordinating on: what pilots need, what technical blockers exist, what's next priority

• Direct access to retail data: We worked with real retailer data (anonymized) from Day 1. Not synthetic datasets—actual transaction logs, inventory feeds, schedule data

• Shared technical infrastructure: FMCG Analytics gave us Azure access, analytics model repos, data schemas. No "submit a ticket to access." We moved at product speed.

Law 3 (Direct Infrastructure Access): We weren't building at arm's length. We worked inside RWA's technical environment. When a model wasn't scaling, we could profile and optimize directly.

Layered Platform with Proven Foundations

Phase 1: Data Infrastructure (Months 1-4)

Objective: Secure, scalable data ingestion and storage.

Critical architectural decision: Don't build data warehouse from scratch.

Modern SaaS platforms need:

• Multi-tenant data isolation (Retailer A can't see Retailer B's data)

• Scalable storage (handle growing data volumes without rewriting infrastructure)

• Fast queries (analytics run quickly even on large datasets)

• Security & compliance (retail data is sensitive—PCI for transactions, privacy regulations for customer data)

This is solved infrastructure. Cloud data warehouses (Snowflake, BigQuery, Redshift) handle this. Building from scratch would take 6-9 months and introduce risk.

Our approach:

• Deployed Snowflake as data warehouse (multi-tenant isolation built-in, scales automatically)

• Built custom ETL layer for retail-specific data sources (POS systems, inventory management, scheduling tools)

• Focused custom development on: data quality validation, retail-specific transformations, client onboarding automation

Time saved: At least 4-6 months. We didn't spend Q1-Q2 building warehouse infrastructure. We spent it building the retail-specific integration layer while the warehouse "just worked."

Law 10 (Avoid Multi-Front Wars): Use proven cloud infrastructure for the commodity layer. Fight the battle that creates competitive advantage—retail data integration and analytics—not the one that's already been won by cloud providers.

Phase 2: Analytics Operationalisation (Months 3-8)

Objective: Turn FMCG Analytics's IP (analytical models) into production software.

The IP operationalization challenge:

FMCG Analytics had valuable analytics:

• Inventory optimization: Predict optimal stock levels per SKU to minimize waste while avoiding stockouts

• Labor cost forecasting: Predict staffing needs based on foot traffic patterns, minimize overstaffing

• Demand prediction: Forecast product demand by category, location, season, promotions

But these existed as Python notebooks and Excel models. To become a SaaS platform, they needed to be:

• Automated: Run on schedule without manual intervention

• Scalable: Work for 1 retailer with 5 stores and 1 retailer with 500 stores

• Production-grade: Handle edge cases, fail gracefully, produce consistent results

Operationalisation work:

• Rewrote models as production Python code (proper error handling, logging, testing)

• Deployed models as microservices (containerized, auto-scaling, monitored)

• Built orchestration layer (schedule model runs, manage dependencies, handle failures)

• Created feedback loops (track model accuracy, flag when predictions diverge from reality)

Example: Inventory optimization model transformation

Before (consulting): Data scientist loads retailer's 6 months of sales data into Jupyter notebook. Runs model locally. Generates recommendations. Emails PDF report to client. Timeline: 2-3 days per retailer.

After (SaaS): Retailer's POS data syncs automatically. Model runs nightly. Recommendations appear in dashboard next morning. No manual intervention. Timeline: Automated, 24/7.

Phase 3: Self-Service UI (Months 6-12)

Objective: Retailers can use the platform without RWA hand-holding.

Core features:

• Data connection wizard: Guide retailers through connecting POS, inventory, scheduling systems

• Analytics dashboard: Visual display of insights (inventory health, labor efficiency, demand forecasts)

• Custom reports: Retailers build their own views, filter by store/category/timeframe

• Alert system: Automated notifications when metrics exceed thresholds (e.g., "Store 7 overstaffed by 15% this week")

Critical UX principle: FMCG brands aren't data scientists. Dashboard shows "what to do" not "here's raw data, figure it out yourself."

Example: Instead of showing correlation coefficients and model outputs, show: "Reduce SKU #1234 stock by 15 units at Store 3. Predicted savings: $840/month."