I ship AI systems that pay for themselves. Not slide decks.

The person responsible for maintenance (tracking, reporting, staying on top of everything) resigned, leaving a lot of loose operational threads with no owner.

I built a maintenance request-tracking PWA (Next.js, FileMaker Data API, Supabase) with generative AI wired into four real friction points: photo-based damage classification, voice-note transcription and cleanup, structured close-out report extraction, and quote/PDF data extraction. All AI calls use structured outputs (JSON Schema) to guarantee parseable results, with a human reviewing and confirming every output. No autonomous agent, no memory between calls: targeted AI inside a deterministic app that also handles routing, notifications, and sync.

The factory has run its maintenance workflow for two months in production without backfilling the vacant coordination role, avoiding roughly $13,000/yr in salary cost, about 13× ROI on what I charged to build it.

The operation delivering bank cards had no real-time visibility into which shipments were stuck. Someone had to open every delivery point's daily Excel ("Punto CLON") by hand and scan colour-coded pivot sheets for trouble.

I built a full operational web panel (not an Excel sheet or a no-code dashboard) that answers, live: is the operation healthy today? Which shipments need follow-up right now? Which points are worst? A shipment is flagged at risk automatically when it goes 2+ days with no status change, or never left for delivery.

Replaces the daily manual Excel review per delivery point. Access is segmented by security (Row Level Security) instead of emailing the whole sheet to everyone. In real daily use.

Couriers delivering cards for four different banking clients (Banco Pacífico, Produbanco, Banco Bolivariano, Cooperativa JEP) had to memorise a long physical manual with different rules per bank: which documents to request, which signatures, which critical warnings.

A step-by-step assistant, built as one self-contained HTML/CSS/JS file with no external libraries, that guides the courier through 1–3 simple questions and produces the exact checklist: documents, required signatures, warnings (e.g. "deliver to the account holder only"), and what to upload to internal systems. All business logic came from the company's real manual, nothing invented. Where the original manual was ambiguous, the output says "verify in manual" instead of guessing.

Validated against all 18 possible rule combinations. Now in active use by couriers across all 4 banking accounts and all 39 delivery locations.

A person manually cleaned Excel sheets before they entered the downstream process: repetitive, slow, and error-prone.

I automated the whole cleanup with a VBA macro.

Saves roughly 80 hours a month of one person's manual work.