I work embedded inside a company for a defined period of time. The engagement is scoped around a specific integration or workflow, starts with the data layer, and ends with agents running in production.
Phase 1 - Weeks 1 to 2
Understand the system. Map the data.
Before anything gets built, I need to understand what data exists, where it lives, and what decisions it is supposed to inform. I work directly inside your stack - reading your APIs, your database schema, your existing integrations - and map everything into a unified data model.
This is not requirements gathering. I am looking for the joins that do not exist yet - the places where Toast does not know what R365 knows, where your CRM does not know what your support data knows. That gap is where the leverage is.
Phase 2 - Weeks 2 to 4
Build the integration layer.
I connect your data sources into a coherent model that an agent can reason over. That means writing the connectors, normalizing the schema, handling the edge cases in your data, and making the whole thing reliable enough to run unsupervised.
This layer is the hard part, and it is where most AI projects fail. A model given bad context produces bad output. Getting the context right is an engineering problem, not a prompt engineering problem.
Phase 3 - Weeks 4 to 6
Deploy the agents. Ship to production.
With a solid data layer under them, agents are straightforward. I build agents that own specific workflows end-to-end - monitoring for trigger conditions, executing the decision logic, taking action within defined bounds, and escalating the exceptions that genuinely need a human.
The goal is a system that is live, running, and measurably working before the engagement ends. Not a handoff document. Not a prototype. Something your team can rely on from day one.
What makes it work
Access to the real system
Not documentation. Not a sandbox. I need to work inside the actual stack - read access to your data sources, your APIs, your schema. The integration layer cannot be built in a vacuum.
One clear workflow to start
The first engagement is scoped to a single workflow. One high-frequency process, one data integration, one agent. That constraint keeps the timeline tight and ensures something working ships at the end. Scope creep is the fastest way to end up with a project instead of a product.
A decision-maker in the room
I move fast and I ask direct questions. When something is unclear, I need someone who can answer it or make the call. Slow feedback loops are the main thing that stretches timelines.