Why I Don't Start AI Projects With AI
A powerful technology isn't automatically the right starting point. Start with the outcome, understand the system, find the constraint, then choose the intervention.
Louis Scalise · AI implementation and systems engineering
More than 15 years in sales, marketing, finance and business operations, then formal study in AI engineering. My work sits where the two meet.
My background spans more than 15 years across sales, marketing, finance, customer acquisition and business operations. That work taught me something technology alone can't: a technically impressive system is useless if it doesn't work inside the reality of the business.
I later studied AI engineering at Maestro College, bringing software, data, automation and AI into the business experience I already had. Today my work lives at that intersection.
I want to understand two things: what the business needs to accomplish, and how to engineer the system that accomplishes it. That means understanding the people who use it, the process around it, the information it depends on, the technology it connects to, the constraints on it, and the evidence that will tell us whether it worked.
I don't believe every problem needs AI. I do believe every implementation needs clear thinking.
My background didn't begin in a computer science classroom. It began with customers, sales teams, operations, financial products and dealerships.
AI engineering gave me the ability to move from understanding operational problems to designing and building the systems that address them.
That combination shapes how I approach implementation. I don't want to build technology that looks impressive in isolation. I want to build systems that survive contact with the business: bad data, missing information, permissions, exceptions and human behavior.
Python, PostgreSQL and DuckDB for data. PM4Py for process mining. Presidio for scrubbing personal data. TypeScript, Bun and Node for services. REST APIs, webhooks and ADF lead files for integration. Activepieces for automation, Metabase for measurement, Railway for hosting. Language models where they earn their place, with a person reviewing what matters.
I spent years working inside sales, customer acquisition, finance and operations before I started engineering the systems around them. That experience changed what I pay attention to.
I'm interested in the places where the documented process and actual behavior stop matching: the handoff everyone assumes happened, the field nobody owns, the automation that technically ran, the exception that disappeared inside a normal workflow.
Those are often small failures. They can also be the part of the system that matters most.
The problems I like are the ones everybody already half-knows about: the leads that "always" get followed up, the report everyone trusts until someone asks where the number came from. Nobody's lazy and nobody's lying. The work just slips somewhere between the people and the tools. I like finding that exact spot, proving it with the records, and fixing the smallest thing that makes it stop.
These aren't methodologies I'm selling. They're ideas I've studied that change the decisions I make.
Technology changes quickly. The principles for reasoning about systems last longer. This is where I write down what I'm learning, what I'm testing, and when experience changes my mind.
A powerful technology isn't automatically the right starting point. Start with the outcome, understand the system, find the constraint, then choose the intervention.
A booking form and a server both worked, and real bookings still vanished. The failure lived on the arrow between them, and it changed how I test.
A workflow is a sequence. A system is everything that sequence needs to succeed: people, information, timing, ownership, exceptions and what happens downstream. Optimizing the workflow alone can just move the problem.
An MVP shouldn't mean a bad version built quickly. Find the assumption carrying the most uncertainty and build something credible that tests it.