VeraConnect Marketing Pipeline

The brief

Before I started, VeraConnect's growth team ran the marketing loop by hand: pore over user activity, decide what to send, draft the email, schedule it, wait to see if it moved the needle. The loop was slow, the quality varied depending on who picked it up, and the only place to learn from a campaign was after the results came back, days later.

The internship question was whether an agent could close that loop. The interesting bit was that this pipeline was for VeraConnect itself — its own user base, its own KPIs, its own brand voice. Not a product I'd be reselling to clients.

What I built

A closed-loop agent system, owned end-to-end inside VeraConnect's platform:

• Ingestion. Pulled user activity from the platform's event store and turned it into a per-account timeline the agent could reason over.
• Reasoning. A planner agent that reads the timeline plus the KPI we cared about and proposes a strategy — segments, messaging angles, send cadence.
• Generation. A copy agent that turns the strategy into concrete email variants, with subject lines and body copy tuned per segment.
• Execution. Wired into VeraConnect's existing email service provider so the planner can actually push the send button.
• Feedback. Reads back its own send metrics and feeds them into the next round of strategy. The closed loop is what moves the KPI; an agent that fires and forgets is just a slow mailchimp.

What was hard

The hardest part was not the orchestration. It was drawing a clean line between what the planner was allowed to decide and what a human still had to approve. I went through three versions of that contract. The one that stuck framed the planner as a recommender with confidence scores: high-confidence runs go out automatically, low-confidence runs wait for a human, sensitive accounts always wait.

The second-hardest part was making the agent inspectable. Growth teams do not trust a black box that "did something." Every decision the pipeline makes is logged with the inputs, the alternatives it considered, and the reasoning trace. A human can audit any individual send in under a minute.

What I took from it

Marketing is a much better proving ground for agent systems than software engineering, because the loop is faster and the metric is unambiguous. A campaign runs in days; you find out whether the agent worked. Code refactors run in months; nobody knows. If you are building an agent system and you want to know whether the architecture is right, start with marketing-shaped problems.

The other lesson was about scoping a summer. Three months is enough to ship something a real team uses daily — but only if you spend the first two weeks watching the team use the current tool before writing a line of code. I did, and it saved me from building something that would have been technically neat but operationally dead.