Learning & Practice
Selective schools, camps and real-world challenges where analytical thinking, AI systems and leadership become visible through the work.
01 · Genesis Academy
Five selected cases from a practical product analytics program: statistics, experiments, funnel diagnostics, marketing analysis and allocation logic.
Pricing · Experiment audit
The weekly price went 6.99 → 9.99 on an A/B test that showed conversion holding, and the revenue never appeared. I re-ran the test on the metrics nobody had checked.
Conversion did hold. The EU upsell fell by a third and LTV rose 2.5% instead of the promised 42.9% — retention was the leak. Roll EU back, keep 9.99 in the USA.
Funnel diagnostics
A conversion drop with no obvious cause. The work was separating a real break in the funnel from an anomaly that would correct itself.
The drop localised to one segment rather than the funnel as a whole, which is the difference between fixing something and waiting it out.
Marketing analytics · Anti-fraud
96M activity rows, 8M registrations and 2.3M payments, searched for partner fraud. Built on DuckDB so the whole analysis re-runs from raw files in one pass.
Two networks named with the evidence: one returning ROAS 0.85 sixty days after registration, one at a 21.24% chargeback rate. Shipped as a Power BI dashboard and a rule set, not a one-off memo.
OBRIOMarketing analytics
How to call a creative early instead of waiting for $2,000 of spend. The honest finding was that a single creative cannot be predicted — so the framework answers a different question.
Not “which one wins” but “how much more, and when to look again”: nothing judged under $50, three decision zones at $200, a hard stop at $500 — because even past $2,000 the verdict still flips 23.8% of the time.
Decision systems
Pick which offer a click goes to, in under half a second, for the most expected revenue per request. The data would not carry the obvious approach, so most of the work went into making the evaluation honest.
A frozen model, scored once on a locked cohort of 127,626 requests, cut prediction error 10% against the next best. No cost data in the set, so the deliverable is a designed A/B — not a promised uplift.
Analytics Camp3.0Select a caseClick a company to inspect the case
02 · KSE Agentic AI Summer School 2026
A selective nine-week program for building autonomous agents against real problems—from LLM foundations and MCP to orchestration, evaluation and production.
Build the context layer first, then connect models to knowledge, tools and external systems without losing control of the task.
Design the runtime around orchestration, state and explicit human checkpoints instead of treating an agent as one oversized prompt.
Move from one-off generation to systems that can inspect their work, learn from feedback and improve through controlled loops.
Make the prototype observable, testable and safe enough to defend on Demo Day and continue hardening beyond it.
×
03 · Case Challenge 2.0 · Team captain
As team captain, I led a seven-day case to choose the strategy most likely to maximize a large e-commerce retailer’s exit value over 24 months.
A scenario-based recommendation and a concrete leadership debrief after we did not advance to the semifinal.
