Applying AI to real business questions.

I started with Backroom, a proposal for an AI operating assistant that compares a small service business with a standard operating model for its trade and tells the owner which functions no one is covering. As I worked through the market and the competition, I made a deliberate choice to serve scientists first. In a lab, the scientist is the knowledge base, and turning that knowledge into computation usually means hiring someone expensive.

That led to Methods, a guided conversation that captures how a research team actually works and turns it into standardized documentation, workflow maps and templates, while the scientist stays in charge of the science. I then sat down with working scientists, watched how they used it, and turned what they asked for into the product requirements.

My instructor called the proposal the strongest reasoning in the cohort, and the idea became Value Stream Science, which is now live. What I took from it is that the best AI products start with a person who knows the work and simply doesn't have the time or the team to organize it. The process mapping and swim-lane analysis I learned in the course are now built directly into the product.

I worked with more than 100,000 public reviews across restaurants, home services and hotels, and treated each sentence as a piece of evidence. Using topic modeling, I built what I called a Category Standard: the handful of things customers in a given category actually judge businesses on.

To score individual businesses, I compared classic machine learning with a fine-tuned transformer model for sentiment. I then used a statistical method that pulls a business with thin evidence toward its category, so every score comes with an honest range. Finally, I used small language models running locally to explain each result in plain language, with code checking every sentence they wrote.

Instead of guessing, I measured how much evidence is enough. A business needs about 14 reviews before its overall score can be trusted. Only a third of restaurants reach that, and just 3 percent of home-services businesses do. So the answer is to read the category, not the business: every business can learn what its customers value, and its own score appears only when the evidence supports it. I also found that in home services, customers care about professionalism more than anything else.

I researched and fact-checked the market, recording a source and an as-of date for every number. Then I built the story in Tableau across five views, showing how the market grew, where it deviated from expectations, how players were distributed, where the money was flowing geographically, and who was backing whom. An interactive dashboard let the viewer filter everything by region.

I closed with a ten-minute recorded briefing written for an executive audience.

My recommendation was for Runway to compete on creative control rather than price, leave the cheapest cost per second to others, focus on enterprise and European customers who value trust and IP safety, and lock in compute partnerships to protect its economics. The bigger lesson for me was that a good visualization is an argument, and every chart should move the audience one step closer to a decision.

I worked with a team of five to write a case brief for the company's leadership. We assessed the company's maturity against COBIT 2019, mapped each failure to the relevant ISO/IEC 27001 and 27002 controls, and estimated the annual loss exposure for each risk.

We then ranked the possible investments by net present value, risk and urgency, so leadership could see not just what to do, but in what order and why.

We recommended four control investments, a hybrid data-governance model and an ISO 27001 management system: about $1.76 million in first-year funding against $7.3 million in annual loss exposure, with a five-year net present value of $14.4 million. It reinforced something I believe about governance: when you put risk in financial terms, it stops being a technical conversation and becomes a business decision.