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
- Designing an OLAP-first analytics pipeline for high-volume, multi-dimensional reporting
- Custom async routing architecture: evaluated LangGraph, selected a purpose-built FastAPI async router for lower latency at production scale
- LLM-as-Judge evaluation framework and human-in-the-loop output quality pipeline
- Cost-governance architecture: semantic caching, token budgeting, dynamic model routing - achieving 65% inference cost reduction ($8 → $2.50/request)
- FastAPI service separation between analytics, application logic, and AI inference
- Scaled to support 10,000 concurrent users with sub-100ms latency targets
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.
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
- Large-scale geospatial feature engineering (200M+ features) across census, mobility, traffic, and infrastructure datasets using Spark on Databricks
- Dual trade-area framework (drive-time isochrones + radial distances) for real-world accessibility modelling
- Cold-start modelling via clustering-based statistical twins for new-to-industry sites
- Multi-vertical XGBoost forecasting (fuel, diesel, prepared food, grocery) with SHAP explainability
- Migration from schema-on-read Hive to star-schema Delta + Unity Catalog: query latency from 45s → <2s (22× improvement)
- Production deployment using MLflow, Delta Lake, and serverless model serving
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.
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
- Gamma-Poisson Bayesian inference: site-level enrollment modelled as a Poisson process with Gamma-distributed rate - enabling native probabilistic uncertainty quantification from sparse monthly data (1–5 patients/site/month)
- Hierarchical MCMC prior elicitation: country × TA × phase × indication distribution fitting with fallback cascade for cold-start sites (PyMC3 NUTS sampler, 2000 samples, 90% burn-in)
- Bayesian conjugate reforecasting: mid-trial posterior updates using IRT actuals (Gamma(α+k, β+v)) - no model retraining required
- IRT-based patient transfer reconciliation: engineering layer to distinguish true attrition from inter-site patient transfers using patient tracking IDs and dropout reason codes
- AWS data platform: Glue ETL pipelines, 3-zone S3 data lake, MICE imputation, CloudWatch monitoring
- Enterprise integration: flat-file monthly batch delivery to SAP IBP for supply planning; Veeva Vault writeback; MLflow model governance
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.
(Additional case studies covering enterprise demand forecasting, MLOps platform engineering, and LLM pipeline architecture in production will be added.)