Autonomous AI Recruiting Pipeline

This screens job applicants from start to finish on its own. Someone applies, AI reads their résumé against the job, they record a voice interview whenever it suits them, the system scores both what they said and how they said it, then decides who moves forward. Nobody on the staffing team touches a normal application. It runs on real candidates today, built straight into the company's own hiring software.

The cost trick is routing. Every AI call goes to the cheapest model that can handle that particular step, which cut the bill 6.5× and brought screening down to about two cents a candidate.

REI AI Command Center

A dashboard for real estate investors with six AI assistants, each doing one job: tracking the portfolio, finding deals, planning rehabs, chasing maintenance, sorting tax, and waking up cold leads. A seventh agent sits above them, works out who should handle a request, and pulls the answers back together. Demoed live at a national industry summit.

Voice Agent for Field Service Close-Out

When a technician finishes a job, the billing team has to chase them by phone for the handful of details an invoice needs. This calls the tech instead. The agent asks a fixed set of questions, and the answers come back mid-call as structured data, land in the job record in the company's ERP, and flag the ones that need a human.

The hard part was writing into a system that already holds real people's work. The ERP only accepts a whole-note overwrite, and roughly a third of those notes already had a biller's own writing in them. So it reads the note first, prepends to it, and refuses to write at all if that read fails — an automation that can silently destroy someone's work is worse than no automation.

Next on the roadmap: the close-out call is the first piece of a larger program. The anchor build is an invoice agent — it reads the finished job, works out what the customer's contract covers and what is billable, and drafts the invoice on hold so a person still releases every one. The model classifies and names the rule it applied; a deterministic step does all the money math, so no dollar figure is ever produced by a language model. After that comes the dispatch side: turning inbound service email into scheduled appointments, and after-hours coverage that knows who is on call.

Real Estate Investment Platform

A 27-page platform for property investors. It rates any deal on eight weighted factors so every property ends up with one score you can compare, and prices a home by fitting a trend line through recent comparable sales — the reasoning an appraiser does by hand, done in a second. It also spots likely flips and drafts the contract.

TRAIGA Readiness Platform

Texas passed a new AI law and most companies have no idea whether it applies to them. This asks ten questions, weighs the answers, and tells you how exposed you are — plus the three things to fix first and where your industry is most at risk. It turns a compliance worry into a lead: people hand over their details to see their score.

Operations & Automation Audit

A paid review of how the business actually runs day to day, and what to automate first. I mapped the work into a five-stage automated workflow, then costed it three ways — cautious, likely, and best case — so they got an honest range instead of one flattering number. Delivered with an interactive page they could click through.

SeanOS — My Second Brain & Operating System

The system I run my working life on. I dump things in — meeting recordings, articles, voice notes, half-formed ideas — and it files each one, links it to what I already know, and hands me a plan in the morning. Two machines share one set of files: my Mac does the building, and a small always-on server runs the daily routine overnight. A watchdog checks every run and repairs the ones that fail.

It is all plain text files. The dashboards and views on top are throwaway — any app can rebuild them — so nothing important is trapped in one program's format. I can move the whole thing to a different app, or a different AI model, without losing anything.

quill — Meeting Notes That Never Leave My Mac

A Mac app that records my meetings and writes up the notes. It records my microphone and the other person's audio as two separate tracks, which is a simple trick that tells you who said what without any AI guessing at voices. The transcription runs on the laptop itself — about 20 seconds of work per hour of audio — so no recording is ever uploaded anywhere.

I picked the parts by testing, not guessing. Apple's built-in on-device model went first: on a real 12,000-word meeting it credited the wrong person with half the conversation and dropped the details, so note-writing goes to Claude and the local model stays as the offline backup. That is the one thing that leaves the machine, it is text only, and it can only reach a single approved address.

Built on an MIT-licensed capture core (digimata/quill). The meeting app, calendar layer, note pipeline, voice-note pipeline, and vault integration are mine.

Staffing / ATS Platform

A 44-page hiring and HR platform built so several companies can use it side by side without ever seeing each other's data. Recruiters drag candidates between stages on a board, and the same system handles HR records and time tracking. GPT-4o writes the job posts, reads incoming résumés, and ranks who actually fits.