LA Department of Water and Power

Overview

Two projects in nine weeks at LA's $6.5B public utility. An AI chatbot that made 250 staff ~80% faster at finding technical standards, and a revenue recovery system that flagged 206 unmetered ADUs out of 80,000+ work requests, projecting $370K+ per year in unbilled electricity LADWP can now recover.

Photo Gallery

solar power's site visit

intern cohort

stakeholder presentation

visiting ladwp's chief ai officer norma

chatting with ladwp's chief resiliency officer zoraya

site visiting pasadena water quality lab

How it worked

PNBED Chatbot. A SharePoint-integrated chatbot for retrieving LADWP technical standards (ESR, NEM, Specs 104). Power New Business staff used to Ctrl+F through 300+ page manuals or interrupt colleagues. The chatbot returns cited answers (clearance distances, NEM system size caps, etc.) directly inside M365. When deployment hit security gates, I built a parallel Python RAG prototype (OpenAI + ChromaDB + Tesseract OCR) so I could keep iterating. The local prototype outperformed Copilot and ChatGPT on every QA test we ran.

NBIAC ADU Revenue Recovery. LADWP had thousands of New Business Installation Accelerated Construction work requests with missing inspections, meaning ADUs drawing power without being billed. I built a SQL/Excel filter pipeline (ADU keyword detection, attribute mismatches, age >1 year, REQ/HOLD status) that took 80,373 work requests down to 206 high-risk unmetered ADUs (0.26%). Documented the logic and handed the flagged list to the ESR inspection team.

• OCR before embeddings: Tesseract page segmentation on inconsistent scanned manuals raised retrieval accuracy before any vector work.
• 3-query cascade for ADU detection: cross-validated by description, comments, and attributes; consolidated duplicates by project number.
• Two tracks for resilience: production SharePoint bot for adoption + local Python RAG prototype for benchmarking when permissions stalled.

Presentation Video

Project Documents