Same moments every time, including the last one — which is the one nobody else publishes.
Described by what they do — the internal names would tell you nothing.
Seller Central BI Platform
Amazon seller · multi-account catalogue
Profit per product was a guess. Fees, ad spend, returns and storage lived in four systems that never met, and the real margin only showed up on payout day.
Orchestrated with Airflow · scheduled, monitored and re-runnable
Handed over: repository, credentials, orchestration and the written definitions — running, and in their name.
Supply Chain Control Platform
Manufacturing to distribution · stock and receivables
Stock lived in the warehouse system, invoices in the accounting one, and what the factories were shipping arrived as email. Nobody could say what was reserved, what was shipped but unpaid, or what to make next.
Three ingestion pipelines · one of them parses what arrives as an attachment
Handed over: repository, credentials, orchestration and the written definitions — running, and in their name.
Emerging Demand Detection Platform
Consumer products · social signal
Demand was moving before anyone could see it. What was about to sell showed up on social platforms weeks before it showed up in sales, and nobody was reading it as data.
Collection runs on a schedule · the archive is the asset, not the snapshot
Handed over: repository, credentials, orchestration and the written definitions — running, and in their name.
Three domains. The same five moments each time, and the same fifth one: governed definitions, exposed so a machine can answer from them. That repetition is the method.
Not every problem needs a platform.
Sometimes the architecture already exists and what is missing is one model, one pipeline, one diagnosis.
Finding the bottleneck by drawing the queries
Retail banking · consumer credit portfolio
Statements took hours and held the server while they ran, delaying every other team. Nobody could say which query was to blame — so we drew them, and it stopped being a guess.
Governed AI over operational metrics
E-commerce logistics · 100+ facilities, 5 countries
Everyone adopted assistants at once and each returned a different number. The fix was not a better model — it was forcing the answer to come from a definition someone had signed off on.
One definition across a continental network
E-commerce logistics · line haul and facility productivity
Every facility measured productivity its own way, so nothing could be compared. Once one definition held everywhere, a model with SHAP could finally say which levers moved output.
Work done inside client and employer architecture, before Horizon Labs existed. Sectors named, companies not — and mechanisms rather than their internal numbers.