KatelynTensley

Level 10 Leverage · Founder & Operations Lead · Jul 2026 – present

Agent Assist

An AI-powered operations platform behind my subscription support business for real estate agents. Agents get everything an assistant would do. Two virtual assistants handle delivery from the SOP library and member portal, with AI doing the repetitive work.

The problem

Agents lose deals in the back office: CRM databases full of duplicates, inconsistent follow-up, and hours spent on MLS entry and presentations. Hiring an assistant is expensive and hard to manage.

What I built

• Member portal. A self-serve web app on Supabase with Stripe subscriptions, role-based access enforced by row-level security, monthly usage quotas, an admin console with KPIs and threaded messaging, and a Database Health dashboard that shows members their progress.
• Two-way ClickUp sync. Every request creates a ClickUp task. A Supabase Edge Function receives ClickUp webhooks, maps statuses both ways, and uses a sync-state column to prevent update loops.
• AI-driven CRM cleanup. AI dedupes, standardizes, tags, and scores agent databases. KW Command has no open write API, so I designed a browser-automation workflow with human approval on every batch.
• Document pipelines. A Node.js and Playwright script renders a 65-field newsletter template to print-ready PDFs and flags anything left blank.
• SOP library. Six SOPs drafted with AI, then checked by a second AI reviewer reading as a brand-new assistant before release.

How I keep AI output safe

AI drafts, a human verifies. Nothing that touches a client's CRM, MLS, or inbox goes out without a person approving it.

Scheduled AI workflow · browser + document analysis · Sep 2026 – present

Compliance Review Automation

Every transaction an agent submits has to be read by a broker against a checklist: is the settlement statement there, is the disclosure signed, do the commission figures reconcile. It's careful, repetitive work, and it's the bottleneck between a closed deal and a paid agent. I built an automated first pass, with broker sign-off, that runs at two Keller Williams market centers.

How it works

• Read the checklist. Find the submitted stage and record every required, conditional, and optional document. A required document that isn't uploaded is a finding before anything is opened.
• Analyze the documents. Pull only that stage's PDFs and check them against the transaction checklist and Florida disclosure requirements, along with the price, parties, and dates.
• Tag, log, and write up. Tag the file, log one row per review, and write a document-by-document summary so the broker opens each file already knowing what's missing.

Design decisions

• Lopsided on purpose. A file is marked clean only when nothing required is missing, unsigned, or in need of correction. Anything the review can't confirm goes to the broker. A false clean costs a compliance failure; a false flag costs thirty seconds.
• Writes almost nothing. A tag, a log row, and notes. It never approves, rejects, messages an agent, or deletes a file, which is what makes it safe to run unattended against live transactions.
• The queue is the to-do list. Already-reviewed files are skipped, so an interrupted run costs nothing. The next run picks up exactly what was missed.
• Stops instead of improvising. If a login is required or a service is down, it changes nothing and reports what blocked it.

Results

Every transaction submitted at two Keller Williams market centers now gets a first-pass compliance review before a broker ever opens the file — twice a day, unattended, with a logged record of every check.

The backlog it was built to clear went to zero and has stayed there. New submissions are reviewed within half a business day, and brokers open each file already knowing exactly what's missing, so they start from the findings instead of a full read-through.

It also found what no human review could. Within the first month, the log showed nearly every finding traced back to the same two documents. One pattern across every review became two fixes upstream — a checklist change and a training point for agents — and the brokers' review time shifted to real exceptions.

The tool flags; a person decides. Brokers sign off on every file.

AI-assisted marketing workflow · Oct 2026

Event Marketing Automation

Every month, agent services staff at two Keller Williams offices promoted 30+ training classes and team meetings by hand: flyers, captions, registration links, scheduled posts on each office's pages, and text reminders to agents. It took hours, and each office mostly promoted only its own events, so agents missed classes they could have attended at the other office. Working with AI, I set up a workflow that runs a full month of event marketing for both offices from one events spreadsheet.

What it produced in October

• 34 events tracked, each with its own Eventbrite registration link.
• 8 new flyers, filed in a shared Google Drive folder.
• About 40 social posts scheduled in Meta Business Suite: one a week out and one the day before each class.
• Every post published as a collaboration between both offices' Facebook and Instagram pages, so each event reaches both audiences.
• 25 calendar reminders so the partner office's front desk accepts each collaboration invite.
• 27 agent text reminders, split by office and scheduled for the day before each event.

Awareness impact

• Both audiences. Every event now appears on both offices' pages instead of one.
• Three touches per class. A post a week out, a post the day before, and a text the day before.
• One shared calendar. Agents at either office now see the whole training schedule, not just their own office's classes.

Python · pandas · rapidfuzz · Claude API

Database Cleanup Tool

A command-line tool that turns a messy KW Command export into a clean, deduplicated contact list in one pass. It's a rebuild of the database cleanup I do by hand for Agent Assist clients.

The problem

Every agent database collects the same rot: one client entered twice with the phone formatted three ways, "Debbie" in one row and "Deborah" in another, an address updated on one copy and not the other. Most tools miss the near-matches that matter or flag false ones, so agents stop trusting the result.

How it works

• Rules first. Exact phone and email matches and high-confidence name and address matches are resolved with plain code.
• AI for judgment calls. Only the ambiguous pairs, like a nickname with no address to confirm it, go to AI, with a hard cap per run so cost stays predictable and a safe fallback if the AI is unavailable.
• Golden records. Duplicates merge into one record, and a phone or email on one copy fills the gap on the other at no cost.
• Pluggable enrichment. Anything still missing can go to a paid lookup provider through an interface built to swap in a real vendor.

The real run

I ran it on a real 500-contact KW Command export from my team. The first run came back with 0% coverage and zero duplicates, and no error. The export used a two-row header, with section labels like "PRIMARY ADDRESS" above the field names, and spread phones and emails across a dozen typed columns. The parser assumed one header row and one column per field, so it quietly read every contact as blank.

I taught it to detect the label row and merge every phone and email column, then reran it: 3 duplicate clusters merged, 500 contacts down to 497, and all 23 automated tests still passing.