Ship AI to production.

RAG concierge

Problem

Customer operations was the bottleneck. Every incoming question required manual search across scattered docs, past tickets, and dashboards. Latency high, throughput flat, no path to scale with hiring.

Build

RAG concierge end-to-end: document ingestion pipeline, vector retrieval over internal KB, retrieval-grounded LLM responses with inline citations, cost and safety guardrails, and an eval harness tracking answer faithfulness, citation coverage, and drift as content changed.

Outcome

2x operational throughput without headcount. Query latency dropped from minutes of manual search to sub-second retrieval. Still running in production, on-budget, within SLO.

Demand forecasting at retail scale

Problem

Manual and coarse demand forecasts across European retail markets meant overstock in some places and stockouts in others, expensive both ways, with no way to scale the approach across markets and SKUs by hand.

Build

Demand forecasting and inventory pipelines across European markets on VertexAI and Kubeflow, BigQuery as the data layer. Hierarchical time-series models, automated retraining, and drift monitoring so forecasts stay honest as demand shifts.

Outcome

Forecasts running in production across multiple European markets, retrained automatically and monitored for drift, replacing manual estimates with a system the team can trust without babysitting.

Real-time NLP at a national publisher

Problem

One of the UK's largest news publishers needed to classify and route editorial content at scale, in real time. Manual tagging was too slow for the volume and editors had no reliable way to test what worked.

Build

NLP content-classification system over real-time data pipelines handling 100M+ monthly page views, plus an A/B testing framework built to be simple enough that editors actually used it day to day.

Outcome

Classification running in real time across 100M+ monthly page views; the A/B framework became part of the editorial workflow rather than a dashboard nobody opened.