I build decision systems for ambiguous strategic questions.
Former machine learning engineer. MBA candidate at USC Marshall.
I turn messy questions about technology and business into working models. Assumptions you can challenge. Evidence you can inspect. Decisions you can explain.
Describe the business, inspect how its strategy becomes measurable assumptions, and rank US metropolitan markets using Census data and an explicit weighted model. The War Room lets priorities change in real time, compares candidates directly, and uses Monte Carlo sensitivity analysis to show whether an apparent winner remains strong when the assumptions move. Every result belongs to the scenario that produced it.
Live-system schematic — the War Room after its built-in premium-fitness example was run on 14 September 2026. The visible ranking belongs only to that scenario and those weights; it is not a universal market recommendation.
Drawn wider than this screen — scroll the figure sideways, or read it in words below.
What this figure shows, in words
The left rail shows editable assumptions for market size, affluence, growth, customer fit, competition, and cost. Competition carries an entered weight but has no data, so the system redistributes that weight and shows a warning.
The main area states that Austin–Round Rock ranks first for this premium-fitness scenario, 0.3 points ahead of Raleigh–Cary. Dallas–Fort Worth appears third. These are the actual results of the built-in example inspected on 14 September 2026, not a general recommendation.
The selected market shows an attractiveness score of 90.5 and factor contributions from customer fit, market size, affluence, and growth. Cost is scored but unweighted in this scenario.
The raw-evidence area cites US Census ACS 2024 observations and explains that growth compares the 2024 and 2019 vintages, while residential gross rent is only a proxy for operating cost.
Enter a public-company ticker and the system reads management's current priorities from the latest 10-K, retrieves the corresponding SEC XBRL financial history, and tests whether spending and operating results support those claims. Language models identify and categorize the stated priorities; financial metrics, thresholds, alignment scores, contradiction checks, and confidence calculations are produced through deterministic code.
Live-system schematic — the public Microsoft example shown by Narrative vs. Numbers on 14 September 2026. It illustrates the separation between management language and deterministic financial evidence; it is not a recommendation about the company.
Drawn wider than this screen — scroll the figure sideways, or read it in words below.
What this figure shows, in words
The management side quotes Microsoft describing plans to expand data-center locations and server capacity to meet demand for AI services, sourced to the 10-K for the year ended 30 June 2026.
The financial side shows the live example's reported changes: capital expenditure plus 79.6 percent, property and equipment plus 52.7 percent, and capital-expenditure share of revenue plus 12.03 percentage points.
The claim-level example scores 100 because all three signals moved in the implied direction beyond the configured threshold. The live page also states that two other claims in the filing score 48.6, demonstrating that one supported claim is not the whole company judgment.
The language model is limited to extracting a label, category, source quote, and confidence. Alignment, confidence, metrics, thresholds, and contradiction checks are calculated from SEC XBRL facts by deterministic code.
02 / THE STRATEGY LAB
Follow the question.
Competitive moves. Product moats. Business value. Three experiments to think with.
A competitive strategy game about defending an established business or building the challenger. Explore decisions, inspect their consequences, and switch sides. The scenario is a fictional project-management market: an incumbent selling seats against a challenger selling completed work. Four quarterly commitments resolve through a deterministic economic model, so every outcome can be traced back to the rule that produced it.
An investigation into what makes an AI product worth paying for. Build a small alternative, test its limits, and examine the business advantages the prototype leaves unresolved. The first case takes meeting assistants: if a model can summarise a transcript, what is the subscription actually buying? Every output is traced to the transcript lines behind it, and each claim carries its data mode.
A financial workbench for exploring the business performance required to justify a valuation. Connect assumptions to operating requirements and test the contribution an AI investment would need to make. It runs the valuation backwards: pick a target, see which growth and margin combinations satisfy it, translate one of those into required customers and price, and then try to break the case you just built.
Start with the decision, not the technology: what choice is being made, what alternatives exist, and what would cause one option to beat another?
02
Structure the ambiguity
Use AI where interpretation is genuinely required, then expose the structured assumptions it produced so they can be corrected rather than trusted implicitly.
03
Ground it in evidence
Connect the structured question to data, documented assumptions, or explicit economic rules rather than allowing generated language to become evidence.
04
Make the trade-offs inspectable
Put weights, formulas, thresholds, missing information, and alternatives where the reader can see what is driving the answer.
05
Stress-test the answer
Try to break the conclusion by moving the assumptions, reversing the perspective, or searching for the conditions under which the recommendation fails.
04 / A LITTLE CONTEXT
From building models to questioning what they mean.
My background is in machine learning engineering, and I am now pursuing an MBA at USC Marshall. I am interested in the decisions around technology: where it creates value, what makes that value defensible, how organizations should act on it, and what would have to be true for the economics to work.