Experience

Demand forecasting — vending and food service

Global food-service operatorApr 2026 – Present

Daily demand forecasting across machine × product series, spanning smooth, intermittent, erratic and lumpy demand classes.

• Replaced per-series Prophet with a single global LightGBM model across 9,200 series and 3.8M rows — full run from ~3 hours to ~3 minutes (~60×), next-day forecast error down ~15%.
• Settled the formulation over 27 controlled experiments: a log1p target with smearing beat Tweedie hurdle staging and the M5-winning DRFAM architecture.
• Cut the feature set from 25 to 15 with no accuracy loss, showing 11 weather variables added no signal to vending demand.
• Took end-to-end ownership of exploration, EDA, preprocessing and feature engineering — now extending them to a broader vending and market dataset.
• Extended the conversational layer of an LLM-driven canteen menu generator (GPT-4.1) that optimises menus against budget and dietary targets.

• LightGBM
• Prophet
• Snowflake
• Python
• GPT-4.1

Model Context Protocol proof of concept

Internal capability build, taken to a client showcaseJan – Mar 2026

• Designed and delivered an end-to-end MCP proof of concept — local MCP servers on the Python MCP SDK, exposing product-catalogue and live-data tools.
• Drove the same servers from three model backends — Claude Desktop, Gemini 2.5 Flash, and Llama 3.1 running locally via Ollama — demonstrating a vendor-agnostic agentic architecture.
• Authored the technical documentation taken to management review, covering MCP architecture, transport layers and server/client implementation.

• Python MCP SDK
• Claude
• Gemini 2.5 Flash
• Llama 3.1
• Ollama

Generative and agentic AI — product recommendation engine

Global B2B chocolate and cocoa manufacturerApr – Dec 2025

Multi-agent system embedded in Salesforce: reads a customer brief — emails, opportunity notes, attached technical spec sheets — and returns ranked recommendations from a 13,000-SKU catalogue, replacing rounds of manual back-and-forth between sales and R&D. Live since June 2025.

• Worked on the summarization agent in a Retrieval-Augmented Generation (RAG) pipeline, turning unstructured customer briefs into structured product requirements — Azure OpenAI GPT-4o, Azure ML Prompt Flow.
• Built the Microsoft Fabric feature-engineering pipeline end to end — all 15 notebooks, turning raw product data into the golden tables the agents query and the embeddings behind retrieval.
• Partnered with R&D and regional commercial teams, working through their documentation, to establish which data sources, columns and SKU combinations drive correct recommendations.
• Engine reached ~80% recommendation accuracy in its core chocolate and cocoa powder categories, with recommendations adopted into live sales opportunities across four regions.

• Azure OpenAI GPT-4o
• Azure ML Prompt Flow
• Microsoft Fabric
• RAG

Revenue Growth Management analytics

Global health and hygiene CPG manufacturerJan – Mar 2025

Trained on the econometrics and purchase-structure workbenches behind an enterprise pricing and promotion platform — how much volume a price move costs, and which products absorb it.

• Ran a hierarchical mixed-effects elasticity and promo-baseline pipeline end-to-end on ~1M weekly sales records, 2,400 products and 21 retailers.
• Reconciled coefficients against a reference run using Welch t-tests and chi-square checks, confirming the pipeline reproduced signed-off results.
• Traced how retailer-level holiday controls keep seasonality out of the price coefficient — the difference between a usable elasticity and a misleading one.
• Worked through purchase-structure trees and volume transfer matrices: how attribute hierarchies become SKU-level switching and category walk rates.

• Mixed-effects models
• Python
• SQL

Sales-driver attribution

Global spirits manufacturer

Assigned to an attribution build isolating how pricing, promotions, media spend and retail execution move retail sell-through across ~600 products and 50 US states.

• Fused a dozen internal and external sources — pricing, promo, media, retail execution, weather, GDP, unemployment — into one product × state × month panel, resolving quarterly/monthly frequency mismatches along the way.
• Applied VIF, variance thresholding and backward elimination to separate overlapping price/promo/media signals, so spend impact was not misattributed across correlated levers.
• Worked with constrained Ridge/Lasso and sign-constrained SARIMAX models, decomposing volume change into per-driver contributions — the interpretable, commercially-coherent output format the client’s team needed to act on it.

• SARIMAX
• Ridge and Lasso
• Python