Project Portfolio & Impact Record

Team members spent hours searching across dashboards and tools to find metrics and build trust in data. Built an AI decision-support platform unifying metrics, dashboards, and operational docs in one conversational interface with row-level access control. Shipped with modular, LLM-agnostic retrieval, visualization, and answer pipelines. 199 people asked 1,838 questions; 79 recurring users across 16 job families; daily active users peaked at 73.

Team members ask questions across internal documentation and business metrics, but no single interface covered both. Built a conversational assistant that routes between sourced-answer retrieval and Looker analytics, with careful handoffs and corporate-proxy-safe long-running responses. Question classification tuned to avoid misleading chart recommendations. Rolled out with governance-friendly design, letting analytics owners tighten behavior before wider release.

Analyst teams build valuable calculators and models every week, but most live outside the main analytics hub: spreadsheets, notebooks, one-off scripts, and small web apps with inconsistent permissions and no shared discovery. Proposed and own a central git repository where tools are built once to a shared standard, reviewed before merge, and callable by conversational analytics as well as through a UI. Proof of concept: ambient/chill SKU cap optimiser and a stakeholder demo tool, with Claude Code guided scaffolding and migration paths for legacy code.

Large automated warehouses rely on periodic reports to track bulky inventory levels, causing delays in restocking decisions. Built a live dashboard comparing current warehouse stock to sales-forecast targets with user-configurable cover levels. Teams can adjust coverage requirements, see gaps and excess at a glance, and act same-day. View-only interface on top of existing systems; requirements gathered through direct warehouse operations walkthroughs.

Commercial teams knew headline sales for bulk beverage promotions but didn't account for delivery cost per tote, driver costs, or incremental route pressure. Compared baseline and promotion weeks for water and soda across warehouse facilities, modeling delivery costs, tote fill from bulky volume, and opportunity cost. Presented methodology and findings directly to client leadership with Q&A; built reusable guardrails for future promotions. Estimated $420K impact at largest site over seven days; $560K across two sites; approximately $1M network-scale exposure.

Analytics team built similar data pipelines repeatedly: moving data between Looker, Slack, BigQuery, and email, scheduling jobs, managing credentials. Created a shared Python framework with reusable helpers for common analysis patterns, automated GitLab releases, secure credential handling, and job scheduling. Established patterns that team members adopted and extended. Grew to approximately 20,000 lines across five production Cloud Functions deployed on Google Cloud.

Grocery partner operations teams couldn't tell which fresh products and locations had excessive pick and delivery costs relative to margin. Built a model identifying where fresh groceries cost too much to fulfill, with clear recommendations. Provided client-side self-service capability so their team could rerun it. Presented findings and model logic to partner operations leadership to drive adoption across the network.

Partner warehouse network received bottled water without capacity limits, causing congestion and inefficient routing. Analyzed delivery patterns and warehouse receiving limits across all locations, surfacing better receiving caps and delivery routes. Presented findings and recommendations to partner operations leadership. Operations team adopted recommended caps and route changes.

Delivery operations team relied on manual rules of thumb to forecast how many routes would run, but needed interim predictions between official planning cycles. Built a forecasting model from historical patterns using ARIMA and scikit-learn, deployed as an automated Google Cloud Function. Model more accurate than previous heuristics. Conducted walkthroughs with partner teams so forecasts informed staffing decisions with confidence.

Partner teams struggled to navigate a large Looker reporting site serving multiple partners and 100+ sites, resulting in low and inconsistent usage. Redesigned home pages and navigation menus based on actual usage patterns, prioritizing the most-used operational reports. Worked directly with partner stakeholders to ensure navigation reflected how they searched for insights. Included self-serve access to critical operational dashboards.

Warehouse managers needed to plan staffing levels but relied on static schedules without regard for demand variation. Built machine learning models predicting warehouse workload and staffing requirements across partner distribution centers. Models deployed to production and integrated into partner operations planning. Covers approximately 15 facilities across 10 partner organizations.

Delivery operations teams couldn't tell when routes could open or whether delivery zones were properly sized based on real customer demand. Analyzed actual booking behavior against operational assumptions, showing patterns where zones were too large or poorly shaped. Built charts highlighting demand patterns and problem areas. Presented findings to leadership to clarify area-setup decisions and enable confidence in automated route-release logic.

Warehouse operations teams need same-day visibility into order volume and labor implications for staffing. Built a live dashboard showing order breakdown and labor requirements, shaped around what operations teams asked for in walkthroughs. Provides real-time performance visibility for day-to-day monitoring and staffing adjustments.

Delivery operations teams had booking data but couldn't visualize patterns and problem areas on a map. Built interactive maps showing booking patterns with filters for time period, site, and slot-planning region. Maps highlighted oversized delivery zones and areas where demand patterns did not match zone boundaries, making problems visible to operations leaders.