An agentic GTM pipeline that runs nightly, built and maintained by a marketing executive rather than an engineering team.
Every night, this site crawls open GTM roles across 11 applicant tracking systems at pre-IPO cybersecurity companies, normalizes and classifies them, scores them against a taxonomy, snapshots the result for historical trend data, and sends a filtered digest to subscribers. The same sequence that governs my first 90 days in a marketing role governs the pipeline: collection is Observe, normalization and scoring are Analyze, digest and outreach are Execute. I designed, coded and maintain all of it, on shared hosting, alone.
Each stage hands a structured artifact to the next. Nothing in the chain depends on a human being awake.
Polls 11 applicant tracking systems nightly across a tracked universe of pre-IPO cybersecurity companies. Streams heartbeats so a long run cannot be killed by a gateway timeout.
Watches a curated set of domains for funding, leadership change and category-launch signals, then writes a structured handoff for the stages downstream.
Deduplicates companies and roles, resolves the same company appearing under several names, and classifies each role by function, seniority and stage using a union-gate taxonomy.
Writes a dated snapshot of the normalized dataset so the hiring landscape can show change over time rather than only the current state.
Builds a per-subscriber email filtered to their criteria, drafts outreach where a signal warrants it, and syncs the result outward to a spreadsheet and CRM layer.
The interesting constraint is not the code, it is the environment. This runs on shared hosting: no container, no queue, no build server, a PHP process that can be killed for taking too long, and an aggressive cache that will happily serve a file you replaced ten minutes ago. Every architectural decision follows from that.
Long jobs stream output so they are never mistaken for hung processes. Data files are flat JSON rather than a database, because a database connection is one more thing to lose at 3am. Every asset URL carries a modification-time query string, because the alternative is shipping a fix that nobody sees for a day. The nav renders from one file loaded through a timestamped loader, so a link change deploys once instead of across nineteen pages.
None of that is clever. It is the same discipline as running a marketing stack on a budget: know the constraint, design for it, and instrument everything so failure is visible before someone else finds it.
AI does classification support, scoring assistance and first-draft copy inside the pipeline. It does not own the taxonomy, the thresholds, the decision about what counts as a signal, or the pipeline itself. That distinction matters, and it is the same one I would defend in an operating role: AI assists underneath, a human owns the number.
The version of this I ran in a full-time seat was a four-agent agentic SDR pilot that produced 26 new enterprise meetings in six weeks for a 36x pipeline-to-spend return. Same principle, different surface: the agents did the volume work, the humans owned the judgment and the pipeline.
Because the argument I would make in an interview is easier to make by showing it. A marketing leader who can specify a system precisely enough to build it can specify it precisely enough to brief an engineering team, scope a vendor honestly, and tell when an agency is guessing. The tools on this site are not a side project. They are the same work, at a scale I can run alone.
I did. Every tool, the nightly crawler, the normalizer, the scoring layer, the digest system and the sync layer were designed, coded and are maintained by one person, on shared hosting, without an engineering team.
The architecture, taxonomy, scoring rules and every decision about what the system should do are mine. AI assists with classification, scoring support and first-draft copy. It is a component in the system, not the author of it.
Because it is the same instinct that builds a demand engine: define the number, model the funnel backward, instrument every stage, automate what does not need judgment.
Not publicly, since it runs live against production data. I am happy to walk through the architecture, the taxonomy decisions and the failure modes in a conversation.