Mostly Python and SQL, with low-code tooling where it earns its place.
Currently at Contractor Independence, working on lead ingestion, verification and routing for US home-services companies. Most of what follows is about knowing when a problem has outgrown the tool it started in.
Replaced a 297-module visual automation, which called a language model 33 separate ways, with a rules-first Python service and a 40-module dispatcher. Several hundred deterministic conditions decide what can be decided from a transcript; the model handles only the remainder. Around 1,000 leads a weekday.
An API key that could create charges and read stored card details was needed by a pipeline running in an environment the whole team and its contractors could read. Split across two instances behind a minimal internal API that returns three fields and nothing else, so a compromised endpoint yields a name rather than billing data.
Inherited a 127-module call tracking scenario, nine levels deep, that was treating each leg of a call as a separate conversation and transcribing every one twice. Audited it end to end, fixed the duplication and the malformed recording URLs, and corrected reporting figures that had been double-counting for months.
A 24-hour hold before outbound contact, absorbing the window in which upstream systems correct themselves. Hold state lives in the CRM where staff can see it; an hourly reconciliation sweep back-fills any lead whose ledger write failed, so a partial write cannot strand someone in the queue.
Figures are measured rather than estimated. Each write-up describes the method behind its number, and what I would do differently.