Livia

Inspiration

We started from a very simple frustration: every condominium resident gets a balance sheet once a year, and almost nobody can actually read it. The administrator may not be stealing. But the PDF is definitely helping nobody. Italian condominium administration is a huge, old, opaque market: 1.2 million buildings, non-standard documents, supplier relationships nobody benchmarks, and assembly meetings where residents approve costs they do not understand.

-> What if every resident could upload the building’s PDF and instantly get the evidence, benchmarks, and next questions they need before approving another year of costs?

What we built

Livia is a resident-facing AI watchdog for condominium balance sheets. Upload the annual PDF and Livia turns it into a structured building record: costs, suppliers, cash, debts, compliance items, anomalies, and a health score residents can understand. The demo runs on real Milan condominium documents. For Via Marco Greppi 7, Livia surfaces issues like €22,243 owed to suppliers against only €2,204 cash, lift maintenance running 126% over budget, and repeated admin charges that should be automatable. It does not just show red flags; it explains the evidence and proposes what to do next. Residents can ask Livia questions, generate a factual anonymous email, or let Livia call the administrator/vendor to request invoices, rationale, and corrective action. The key loop is: PDF in, red flags out, resident-approved action taken. Under the hood, we built a multi-stage pipeline: PDF extraction, strict schema parsing, arithmetic validation, anomaly detection, benchmark normalization, health scoring, and verified cached fixtures for demo-critical numbers. The LLM explains and extracts; the system validates, reconciles, scores, and gates.

What we learned

The hardest part is not making an AI summarize a PDF. The hard part is making it trustworthy when the PDF is messy, inconsistent, and financially consequential.

Challenges

Our biggest challenge was balancing demo magic with financial reliability. A live model can produce impressive explanations, but the numbers on stage have to be deterministic.

Built With

• elevenlabs
• fastapi
• mapbox
• openai
• pdfplumber
• python
• react
• resend
• twilio
• typescript
• vite