Case Studies

Three production deployments, built and delivered inside enterprise environments alongside customer leadership. Each system is in live operation. Each directly shaped a measurable, irreversible business decision - capital allocation, drug supply planning, or operational strategy at Fortune 500 scale.

Each case study documents the discovery process, architectural decisions made under real constraints, alternatives evaluated and rejected, and the business outcome - not what was built, but what changed because of it.


Case Study 1: Multi-Agent GenAI Analytics Platform

Client Engagement: Large-scale eCommerce client
Role: Principal Architect
Impact: $100K/yr saved · 40 FTE automated · 65% inference cost reduction · 1,000+ concurrent users

Objective
Design and deploy a scalable, cost-controlled multi-agent GenAI platform to automate weekly and monthly executive business reviews for category managers and senior leadership, replacing manual, error-prone analysis workflows operating on TB-scale transactional data.

What this case study covers

Design Philosophy
Built as a batch-oriented, enterprise-grade analytics platform prioritising consistency, explainability, and operational reliability over real-time inference. The LLM augments human decision-making rather than replacing core business controls.

➡️ Read full case study →


Case Study 2: Geospatial ML - New Store Site Selection & Sales Forecasting

Client Engagement: $10B+ convenience & prepared foods retailer (2,500+ locations)
Role: Lead Architect
Impact: $50M+/month CAPEX informed · 70% faster approval cycles (3 weeks → 5 days) · 22× query latency improvement · 15% accuracy lift vs 3rd-party tool

Objective
Design and deploy a production-grade geospatial ML system to support new-store site selection, replacing decentralised, intuition-driven real estate decisions with a standardised, explainable, and scalable forecasting engine capable of estimating 3-year category-wise sales from a latitude/longitude input.

The system directly informs $50M+ monthly CAPEX allocation decisions, where errors translate into long-term lease risk and irreversible investment commitments.

What this case study covers

Design Philosophy
Built as a spatial-first, cross-sectional modelling platform prioritising interpretability, reproducibility, and operational scalability. Architectural choices ensured the solution could be trusted and adopted by real estate and finance leadership at enterprise scale.

➡️ Read full case study →


Case Study 3: Bayesian MCMC Clinical Trial Enrollment Forecasting - Oncology Supply Chain

Client Engagement: Global Pharmaceutical Client (Top-10 Oncology Biopharma) - Phase 2/3 clinical trials
Role: ML Engineer & Data Engineering Lead · MLOps Architect
Impact: ~$2B projected supply waste avoidable · VP-level sign-off replaced a decade-old 2× heuristic · 63% MAE (first-ever quantitative forecast in client’s oncology history) · 8 trials · 40 countries · 3-year supply plan horizon

Objective
Design and deploy a site-level probabilistic enrollment forecasting system for oncology trials, replacing a blanket 2× over-ordering safety stock heuristic with a Bayesian MCMC model that generates 80% confidence enrollment projections per site, per country, for a 3-year forward horizon - feeding directly into SAP IBP for drug and placebo supply planning.

The problem required solving for patient attrition, site-level data sparsity, blinded trial supply complexity, and cross-site patient transfers tracked via IRT, at a scale of 40 countries and 80+ trial sites simultaneously.

What this case study covers

Design Philosophy
Built as a Bayesian-first, hierarchical probabilistic system where uncertainty is a first-class output. The 80% credible interval - not a point forecast - is the primary deliverable, enabling supply planners to make defensible safety stock decisions without systematic over-ordering.

➡️ Read full case study →


(Additional case studies covering enterprise demand forecasting, MLOps platform engineering, and LLM pipeline architecture in production will be added.)