I move ambiguous AI use cases into audited production systems.

ProblemFares move constantly across dozens of suppliers. Stale and anomalous pricing reached customers as failed bookings, 8% of them, driving refunds, support load and lost trust. No human team can watch 10M+ daily updates across 30K+ routes.

The product decisionI scoped an agentic system that acts, not just alerts: it detects pricing anomalies and autonomously blocks affected inventory within two seconds. The hard call was calibrating autonomy. I defined the detection thresholds, the false-positive guardrails, and a human-in-the-loop review path for low-confidence blocks, so the agent moves at machine speed where it's confident and defers to a person where it isn't. Over-blocking starves inventory; under-blocking hits the customer. That threshold was the product.

OutcomeStale-fare booking failures fell from 8% to 1%, fare jumps stayed under 1%, and the system protected USD 100K+ in annual refund and support cost.

ProblemMarkups were set by flat heuristic rules. Only 60% of route combinations cleared the profitability bar, while others were priced past what customers would pay.

The product decisionI defined the success metric as a business metric, the share of route combinations clearing net profit, so the model optimised against money rather than error rate. We benchmarked a Random Forest Regressor against a CatBoost Regressor, and CatBoost shipped, because the profit lift was large enough to justify the added complexity. That was the real trade-off. A gradient-boosted model is harder to explain to commercial stakeholders who have to trust a price, and heavier to serve and monitor. I took that cost deliberately and owned the work of making the model legible to the business and monitored in production, rather than defaulting to a simpler model for comfort.

OutcomeProfitable route combinations rose from 60% to 85%.

ProblemRoutine travel queries needed deflection at volume. But on a fare rule or change fee, a confidently wrong answer is worse than no answer, so quality couldn't be judged on impressions.

The product decisionI treated quality as an evaluation problem, not a QA afterthought. We built an eval framework with explicit edge-case and adversarial categories, plus fallback logic and human escalation paths for low-confidence turns. That framework became the release gate through UAT and the live monitoring signal afterward, the bot only widened its scope once it cleared the bar.

Outcome95% accuracy in production, with defined human escalation paths.

ProblemLaunching FR, ES, CA, DE and MX meant localising large volumes of customer-facing travel content fast, where a raw MT error becomes a brand or legal problem in a market you've just entered.

The product decisionAn LLM translation pipeline with a human QA gate. The trade-off was explicit: pure human translation is accurate but slow and costly; pure MT is fast but risky. HITL captured most of the speed while holding the accuracy bar. I planned the phased market cutover, rollout, user enablement and training so the pipeline was actually adopted internally rather than bypassed.

OutcomeTurnaround cut 50% at 99% accuracy across all five markets.

ProblemModel training and analytics were bottlenecked on legacy infrastructure. Migrating is the kind of program that quietly destroys a quarter if sequencing or data quality is wrong.

The product decisionRather than a big-bang cutover, I owned estimation, sequencing, schema and SQL conversion, environment and cutover planning, and data-quality validation for a phased migration onto BigQuery, Dataflow and Cloud Composer, moving training and deployment workflows onto Vertex AI, coordinating a 45-person engineering, data science, DevOps and QA group. Phasing cost more calendar time and bought a reversible path at every step.

OutcomeDelivered with no production downtime, a 99.9% pipeline success rate, and a 30% infrastructure cost reduction worth USD 350K annually.