How I decided, not just what happened

The problem

Enter 30+ emerging markets with sparse or absent sales history, 25,000+ SKUs and heavy seasonality. The old approach retrained per series and broke on new products; planners fell back on manual adjustment, inflating safety stock and tying up working capital.

My role & decisions

I built the forecasting & decision-intelligence function from scratch — no inherited team or stack — and owned the architecture and model strategy. I chose a zero-shot, cross-learning model over per-SKU training so new items were forecastable on day one, and pushed to fold external signals (weather, promotions, events) into the feature store. I hired and led the 30-person org and governed EY, Deloitte and BCG delivery; full P&L accountability.

Architecture

MLflow + drift monitoring close the loop back into retraining.

Stack

The problem

Planning knowledge lived in dashboards and people's heads. Answering "why is Southeast Asia short next month?" meant digging; exception triage was manual. And a regulated FMCG org wasn't going to trust a chatbot that hallucinated numbers.

My role & decisions

I led the first large-scale LLM deployment at Unilever. I insisted the system be retrieval-grounded on the planning data with a deterministic validation layer before any figure reached a human — and that a person stayed in the loop for actions. The real work was change management, not the model.

Architecture

Stack

The problem

Enterprises are wiring agents to a single vendor runtime — unobservable, unrecoverable mid-task, and impossible to govern to GDPR / HIPAA / EU AI Act. Fleets of agents fail in ways a single model never does.

My role & decisions

I'm architecting production agentic platforms, including the semantic ontologies and context substrates that let Google Gemini agents reason consistently across heterogeneous enterprise data. My core decisions: decouple workflows from any one runtime (A2A + MCP), and treat governance and self-healing as first-class, not bolted on. This work is in flight; I'll update the outcomes as it ships.

Architecture

State-aware async runtime with checkpointing and self-recovery underpins execution.

Stack

The problem

Diagnosis and procedure detail sat in unstructured clinical notes. Manual ICD-10 coding was slow and inconsistent — and inside a HIPAA environment you can't ship a model you can't audit.

My role & decisions

As lead data scientist / delivery manager I designed the document-intelligence pipeline and the compliance-grade validation around it: cross-validation, AUCPR and full audit trails from the start, not a retrofitted review, with a human coder always able to trace a decision.

Architecture

Stack