shipped

ArborKey

The business layer for community association managers.

Visit site

HOA management companies have unique needs — odd time tracking, their own kind of pipeline and prospects. The platforms they pay for run the communities, not the management business. No big-box software fits exactly. So CappaWork built the one they needed.

It puts pipeline, time, budgets, and reporting in one place. AI does the tedious part: turn a meeting transcript, a board roster, a business card, or a messy CSV into clean records. A human always confirms before anything saves — because in this business, wrong data means wrong bills.

At a glance

Problem

Community association management firms run dozens of HOAs on spreadsheets and sticky notes. The operational platforms they pay for manage the communities, but offer nothing for running the management business itself.

Solution

I built a multi-tenant SaaS for the management company — pipeline, timesheets, budgeting, profitability reporting — with four AI extraction features that turn meetings, board rosters, business cards, and arbitrary CSVs into clean, structured records.

Why it was hard

In a vertical where wrong data drives billing and contracts, AI silently writing the wrong thing is worse than no AI. So every feature is single-turn extraction with a clean → parse → coerce → safe-default pipeline and a human-confirmation invariant encoded in the schema — never an autonomous loop.

Impact

It’s a production multi-tenant SaaS with the full commercial stack — auth, Stripe billing, transactional email, analytics, cron — plus four AI features that kill the transcription busywork, all under a documented philosophy that lets a trust-sensitive vertical actually adopt AI.

AI components

  • LLMs

    A single model — Google Gemini 2.5 Flash — for high-volume, text-in / JSON-out extraction, in JSON response mode on every call.

  • Agents

    None — deliberately single-turn; the “agency” lives with the human reviewer, by design.

  • RAG

    None — bounded zero-shot extraction over self-contained pasted text, where retrieval would add cost and latency for no accuracy gain.

  • Evaluations

    Type-validation plus mandatory human confirmation as the safety gate; precision/recall and acceptance telemetry (PostHog) identified as the measurement gap to close.

  • Workflows

    A uniform clean → parse → validate/coerce → human review → confirm → write pipeline across all four features, with safe-default degradation on a bad generation.

  • Automation

    Stripe billing lifecycle, deduplicated transactional email, and two cron-driven engagement sequences.

Technology

Next.js • Drizzle • PostgreSQL • Gemini • Stripe