AI, customer operations, and operating transformation.

Treated it as a design problem rather than a hiring problem — establishing what had to be true before capacity was allowed to land: work routing, quality ownership, how a new site reaches standard rather than merely opening, and escalation paths across time zones. Stood up an entirely new partner operation in a market the organization had not operated in.

50 to 5,000 agents in one year, including a new site in Cairo.

Mapped the portfolio by risk tier, separating high-trust, high-consequence work from operational queues, then designed governance to match: explicit decision rights, defined escalation, named ownership per tier. Stood up a strategy function that had not existed, connecting frontline operations to enterprise decision-making.

A unified operating model across 50+ lines of business, at roughly $200M operating budget, up to 1,300 internal employees and thousands of partner resources globally.

Defined machine answers as a third collection surface rather than folding them into social listening. Scored three gaps: divergence from the organization's own account, from its ecosystem's, and assertions neither supports. Recorded two lanes per answer — trained belief versus live retrieval with citations. Derived prompts from the organization's own ecosystem, then curated by hand.

324 answers collected in the first run, 400 in the second, across four engines. A companion value model — Decision Exposure, not cost — with a measurement ladder that must resolve to material mismatch before any figure touches money, and a published validation cascade naming which of its own stages remain unproven.

Built extract, score, adversarial-skeptic and synthesize stages, where findings must survive challenge before entering the evidence ledger. Defined a correctness metric weighting reproducibility across repeated trials, evidence verification that checks whether cited identifiers actually exist, diagnostic value by finding type, and cross-run stability. Ran controlled model comparisons on a fixed corpus. Locked a production model standard with an explicit rationale, and a rule that scoring stays on the calibrated model even when extraction moves.

Scoring determinism reached zero variance across identical runs. 73 experiments across four research tracks. A published head-to-head finding that the larger model did not win, and that prompt edits raising acceptance rates broke reproducibility — a tradeoff most teams never measure.