Leary Labanlic

AGM Corporate Library

An AI research assistant for a property portfolio.

Role
Marketing Strategy & Technology Lead
Timeframe
2025
Stack
React · Cloudflare (Pages/Workers/D1/R2) · Vectorize/AutoRAG · Anthropic Claude
Headline metric
Doc search: hours → seconds; ~15 hrs/week saved

The problem

Across 45+ properties, answers lived buried in leases, SOPs, vendor contracts, and financials scattered across email and shared drives. Finding one fact meant opening a dozen documents — slow, repetitive, and easy to get wrong.

The approach

A retrieval-augmented (RAG) document assistant — staff ask in plain English and get answers with citations to the source document. The differentiators are about trust and precision, not just search:

  • Agentic loop (Claude Sonnet) with a 6-tool toolset — the assistant reasons over the corpus rather than doing a single lookup.
  • A second Haiku citation-verification pass — it never cites what it didn't actually read.
  • Focused mode — pin specific documents when you already know where to look.
  • Folder-scoped search — avoids cross-property false positives.

A 15-case golden eval suite guards quality so changes don't quietly regress answer accuracy.

How it's built

React + Vite front end; Cloudflare Pages Functions (edge) + D1 + R2 + Vectorize/AutoRAG; Anthropic Claude with prompt caching. Roughly 8.6K lines of code. It also includes a Reputation Portal — AI-drafted review replies behind a human approval workflow.

The outcome

Document discovery dropped from hours to seconds, reclaiming about 15 staff hours every week. Answers are trustworthy because citations are enforced, and every interaction is captured in a full activity audit trail.

Visuals

Chat answer with inline citations (demo data)
Plain-English answer with citations back to the source document.
Document library with folders + tags (demo data)
The corpus, organized by property with folders and tags.
Focused mode pinning a document (demo data)
Focused mode — pin specific documents to scope the answer.
Reputation portal: draft + approval (demo data)
AI-drafted review replies behind a human approval step.
Activity log (demo data)
Full audit trail of questions, answers, and sources.