Logistics · Forecasting · MLOps
A Canadian logistics scale-up
The problem: Their demand forecasting model looked brilliant in the notebook and useless on the road. Route volatility, weather shocks and a merger that doubled SKU count had not been reflected in the feature set. Planners stopped trusting the output and reverted to spreadsheets.
Our approach: A two-week discovery sprint mapped data pipeline gaps and identified where the old model failed silently. We rebuilt the forecasting model with regime-aware features, shipped a data engineering layer that ingested live dispatch feeds, and stood up model drift monitoring with alerts when error bands widened.
What we built: Machine-learning models for weekly and daily demand, an analytics dashboard for planners, and MLOps pipelines with automated model evaluation. Human-in-the-loop review stayed on every override above a threshold.
What stayed human: Final routing decisions, exception handling during weather events, and quarterly feature reviews with the ops team.
Illustrative outcome: forecast error band narrowed on stable lanes; planners reported higher trust scores in a follow-up survey. Not a guarantee of results for similar engagements.
Retail · Analytics · RAG
An Ontario retailer
The problem: Twelve BI tabs, three conflicting definitions of margin, and a support queue that spiked every Monday because nobody could answer inventory questions without paging finance.
Our approach: We consolidated metrics into a single analytics dashboard wired to the warehouse, documented every definition in plain language, and added an LLM assistant grounded in their data via retrieval-augmented generation (RAG) for ad-hoc questions.
What we built: Reporting layer, data pipeline refreshes, AI assistant with guardrails and citation of source tables. API integration connected to their internal ticketing tool.
What stayed human: Pricing decisions, promotional planning and any assistant response that triggered a stock movement.
Illustrative outcome: median time-to-answer for inventory queries dropped in pilot stores. Adoption varied by team; results are not guaranteed elsewhere.
Margin definitions agreed in the room before the dashboard went live.
Insurance · Data pipelines · Monitoring
A national insurer's ops team
The problem: Claims triage relied on a patchwork of exports, manual joins and a model that had not been retrained since a regulatory change eighteen months prior.
Our approach: Data engineering first — we rebuilt the ingestion path, added quality checks and monitoring on the pipeline itself. Then we retrained the triage model with responsible AI review on protected attributes and shipped model evaluation dashboards the compliance team could read.
What we built: Data pipelines, ML model refresh, monitoring desk with drift alerts, and documentation for PIPEDA-compliant data handling.
What stayed human: Final claim adjudication, appeals and any automated suggestion above a confidence floor.
Illustrative outcome: pipeline failures detected before downstream jobs ran; compliance signed off on evaluation reports. Past performance only.
Monitoring desk — drift alerts configured before production handoff.
Manufacturing · Feature engineering
A Quebec manufacturer
The problem: Sensor data existed but nobody had turned it into a signal. Downtime was investigated after the fact, not predicted with enough lead time to schedule maintenance.
Our approach: Feature engineering on vibration and temperature streams, a proof of concept forecasting model for equipment failure windows, and an analytics dashboard for the maintenance lead.
What we built: Streaming data pipeline, forecasting model with stated accuracy limits, and workflow automation for maintenance tickets — triggered only above a confidence threshold with human approval.
What stayed human: Maintenance scheduling, vendor selection and any shutdown decision.
Illustrative outcome: two unplanned stoppages flagged in pilot window. Scope and data quality drove results; not guaranteed at other sites.
AI & results disclaimer: Our studio provides AI and data design, development and consulting services, including strategy, machine-learning and generative-AI applications, analytics, data engineering, AI assistants, workflow automation and monitoring for client organizations. AI systems can produce errors, biased or inaccurate outputs and require human review and oversight; we design for humans-in-the-loop. We do not guarantee specific business results, predictions, cost savings, revenue, accuracy or return on investment. Outcomes depend on data quality, scope, budget and adoption. Any case studies or metrics shown reflect past client work and are not a promise of future performance. This is a professional AI services firm, not legal, financial, investment or medical advice, and not a course or income opportunity.