The work
I don’t type most of it anymore.
BizPortal is the software that runs Vultun. I design it, read it, and approve it. AI coding agents write nearly all of it. It went live on May 23, 2026.
How it happened
I learned to code at forty, in a boot camp, Googling my errors. I was a data analyst, not a software engineer. In January 2026, I started directing coding agents on the company’s software. In July I wrote: as fast as I can, I’m learning AI vibe-coding with Claude Code.
Then, in the slow season, I cancelled the tools we were renting: Airtable, Zapier, Kit, Cursor, and CopyAI. We got ₱7,854 a month back, and the company owns what replaced them.
It all runs on our own VPS, a rented server at DigitalOcean that never switches off: BizPortal, the customer chatbot, the email engine, and more than sixty scheduled jobs. The database lives at Supabase. The code lives in GitHub, so I can work on any of it from anywhere in the world, and it keeps working 24/7 for every location, whether I’m online or not.
What it does
Our own marketing and email engine
Flows, audiences, and send limits in our own database. It replaced Kit and Zapier. There’s no AI in the sending path, and every send has to pass the gates.
Customer intelligence
A customer feature store, RFM segmentation, engagement scoring across every channel, and six behavioral personas, first explored with k-means clustering.
The customer chatbot
On WhatsApp, Instagram, and Messenger. It reads live class times and a knowledge base we wrote, and it hands off to a person when it should.
- Payroll
- A general ledger
- End-of-day close
- Receipt capture
- Time clock
- Roles and permissions
- Booking and payment sync
Internal only. Staff use it every day. You can’t log in, and the real screens are full of real people. So here’s one with made-up numbers.
In numbers
- 191,412
- Lines of TypeScript in BizPortal. 959 files. 125 test files. Counted November 3, 2026
- 216,601
- Lines of JavaScript in the automation layer: the email engine, the chatbot, the scheduled jobs. Tests included. Counted November 7, 2026
- 15,390
- Lines of Python in customer intelligence. Counted November 7, 2026
- 182
- Database tables, every one behind row-level security. Counted November 3, 2026
- 62
- Scheduled jobs running without me. Counted November 3, 2026
- 83.0%
- Receipt reading accuracy on a local model, with zero invented values, on 48 test receipts. Eval of September 15, 2026
- 88
- Scripted customer questions the chatbot is tested against. 76 passed on the last full run. None tripped a safety gate. Run of November 2, 2026
The rules I run agents by
The AI talks. Functions decide.
- AI interprets and phrases. Functions decide and fetch. An agent can draft a payroll answer. It can’t be the thing that pays anyone.
- Nothing writes to the database without me seeing it first. Every write gets described, with the row count and how to reverse it, and then it waits for my go.
- Check the real thing, never the model’s claim. Merged doesn’t mean applied. A number you can’t verify is “unknown,” never a guess.
- Build the loop, not just the system. Every correction I make gets written down where the next run will read it. The loop is the asset, not the model.
- Autonomy is earned. Money, customers, and health data stay behind a gate that can say no.
Work with me
I can help you build this for your business.
AI automations, and getting AI into the way your company actually runs. It’s one of the biggest things I can help with.