Arnaldo De Lisio

My family's construction company, public-works contractor · Delivered and handed off

• Bidding taken from 1–2 to 10–20 tenders a month in-house, and its first win in the quality-scored category the company could not enter before, a 2 million EUR contract.
• Before this work the company bid on one or two public tenders a month with outside consultants preparing them. It now competes in tenders awarded on the most economically advantageous offer, scored on quality as well as price, which it could not enter before. The 2 million EUR contract won this year was its first of that type.
• I built the multi-agent tender-analysis platform behind that: Gemini extraction of visually hostile scanned bills of quantities and drawing sets, Supabase Postgres with pgvector, hybrid retrieval fusing keyword and embedding passes with Reciprocal Rank Fusion because tender work turns on exact article codes and norm references, and five specialist agents in parallel (safety, cost, compliance, technical, timeline) reviewing 30–50 documents in a few hours instead of a week, with human oversight on the bid decision.
• Beyond tenders, I built the company's automation for accounting, administration and tender participation, and handed all of it over so it runs without me.

Workplace-safety compliance, regulated document production

• 3 hours to about 5 minutes per regulated compliance package, about 400 hours of inspector time freed a year.
• I architected and shipped the production platform that turns a site visit into a compliance-document package: voice, text and photographic evidence captured in the field, written to Postgres with a provenance stamp the moment it is heard, and deterministic Python generators producing the legal documents from that database, so the document is a view of the evidence rather than a thing someone typed.
• Professional review, traceability and accountability are preserved; one inspector at about three inspections a week is where the 400 hours come from.
• Claude Haiku extraction with a Claude Sonnet vision fallback, a custom MCP server inside the Next.js deployment as the scoped tool surface, state arbitration in per-table Postgres functions; transitioned to partner-led operation on self-hosted infrastructure.

Direct-to-consumer luxury fashion, made-to-order production

• About 50% less manual coordination across 8 operators, on a platform built from scratch and in production within 6 months.
• I designed and shipped a complete order-to-cash and production ERP from scratch: customer master data, a thirteen-state order lifecycle from intake through design, pricing, approval, payment, production and delivery, every order carrying explicit owners for customer, designer, tailor and coordinator, billing through Stripe, fulfilment through Shippo, document generation and role-based access across six roles, on one Supabase Postgres model.
• Next.js and TypeScript on the front, a separate Node service on the back; the operators gained shared visibility of every order and who owns it.

Film technology and virtual production

• This was not about building production software; it was about where AI could realistically fit inside professional film production. The question was not what AI can do but what people would actually trust inside their workflow.
• I worked with production and technology stakeholders to find where AI created practical value in creative and operational workflows, and turned that into MVP concepts and integration approaches around existing infrastructure, controls and human decision points. The exciting ideas tended to add more complexity than value; the ones that looked boring removed friction, and those were the ones worth carrying into planning.

B2B marketplace for film and television production resources

• 2,500+ registered users and 120 paying customers, built alone in 6 months, still in daily use after I handed it over.
• Sourcing on a film production is a recurring, high-friction problem: a specific piece of equipment, crew capability, location or service, needed inside a window measured in days, found through a phone book of personal contacts. I had lived on the demand side of it for ten years, so I built the marketplace that replaces the phone book.
• Web, iOS and Android: Next.js, a Node service on Google Cloud Platform, Flutter, Supabase Postgres, Sentry from day one. Handed over to someone I trained, who maintains it without me.

20+ projects, including Men in Black: International, Sense8, The Equalizer and My Brilliant Friend

• Suppliers, schedules, logistics and cross-department dependencies on productions of 300–500 crew with immovable deadlines, directly coordinating daily teams of 10–20: where the ownership, traceability and contingency habits I apply to AI delivery come from.