Built 2026
Opportunity-detection engine
A system that tunes itself on its own evidence
Outcome
Built a system that runs twice a week across seven sources and reads two different kinds of signal — contract and tender postings already shaped like my work, and organizations publicly admitting an operational problem with budget attached. It tiers what it finds, remembers what it has already seen, and surfaces only what needs a human. Runs unattended.
How it was built
Seven independent sources — federal tender open data, a nonprofit job board, two remote-work boards, a community hiring feed, a monthly thread, and one source read purely for market intelligence. Any single source can fail without stopping the run. Feeds are filtered with code rather than read whole, so a several-hundred-kilobyte file costs nothing to scan.
State persists between runs: a ledger de-duplicates by organization and title, so a lead reads as new or still-open rather than arriving fresh every time; a pattern log compares each run against the ones before it; an outreach log marks organizations already contacted. Enrichment is capped per run, with the overflow carried forward rather than dropped.
What it demonstrates
A system that gets tuned by its own evidence. Search categories are promoted or retired on measured yield — several were cut after producing near-zero across repeated runs, and the location rule gained an exception once the pattern log showed it was filtering out the best-fitting segment. That’s the same discipline as the rest of my work — don’t trust a system’s account of itself without an outside check. (It happens to be my own lead-generation tool, which is why I can speak to how it behaves over months rather than minutes.)
Call me when
You have a repetitive monitoring-and-triage task eating time that should run itself and surface only what needs a human.
Contact
If this sounds like a problem you have, the best place to start is a short conversation about what’s actually stuck.
Email me