Building with language models since 2019
Applied AI builder
I learned a construction business from the accounts receivable desk up, then rebuilt how it runs with language models.
2019:
Started building with language models, on GPT-2
110
AI tools in production behind one connector
11 of 11
People with access who use it every week
From about 15 days to 3:
Days to close the month, after rebuilding the stack
$25,000
Yearly software spend removed by connecting systems directly
More than 50
Codex agents running in parallel on one Screed build
01The path
I didn't come to AI from a lab. I learned how a contractor gets paid, buys material and closes its books, then wrote software for work I already knew.
2010
Hand-written HTML, CSS and JavaScript. I've been writing code ever since.
2011–15
Financial services representative at TD Bank. I originated loans as an NMLS-registered loan originator, held life, health and variable annuity licenses, and served as a notary.
2015–19
Finished a B.B.A. at Elmhurst University while running restaurant floors as a server, bartender and supervisor. On the side I built websites and ran marketing for freelance clients.
2019
Started in accounts receivable at a South Florida commercial hardscape subcontractor ($12–15M revenue). The same year GPT-2 went public, and I started building with language models.
2020–24
Procurement specialist, then project manager, controller, and head of finance in 2024. Before LLMs I built Excel tools in VBA and the company's item database in QuickBooks Desktop. I picked and implemented the stack (Procore, Ramp, Rippling, Fleetio, Linxup, Coast) and moved AP off paper checks.
2023
Started my own consultancy for paid software integrations and AI work for outside clients.
2025
Started building the company's AI tools on Gemini, then moved us to Claude and ran the rollout and training myself. It grew to 110 role-scoped tools behind one connector, and 11 of 11 people with access use it every week. Founded Screed the same year.
2026
Rolled out Grok Bot to the team and taught agentic workflows for multi-step work like month-end close. I build Screed by directing 50+ Codex agents in parallel, with review and test gates before anything merges. Started Tellson Press.
02The platform
Staff open Claude or ChatGPT on a laptop or a phone. The company connector signs them in with Google, checks their role, and shows only the tools that role allows. Those tools call Procore, Ramp and QuickBooks Online through their APIs and reach Bluebeam and AutoCAD through MCP.
Feedback → new tool in 1–2 days
03Work
Estimating and operations software for hardscape contractors. It's live with a paying customer that integrates through its API, connects to Ramp and QuickBooks, and uses ambient capture so crews can record work without stopping it.
I build it by directing 50+ Codex agents in parallel behind review and test gates. Evaluation comes first. I rejected a candidate reference dataset when only 4 of 27 checks agreed.
getscreed.comMy consultancy for systems integration and AI work with finance and operations teams. I connect ERP, project management and AP, train the people who run them, and cut software that doesn't earn its cost. I publish results from my own work there.
analysync.comAn independent imprint. Its first title, Entrustment (February 2027), was made with a generative language model under my direction and comes with a public disclosure of what was kept, what was generated and what is unresolved.
tellsonpress.comA GPU workstation running local models through llama.cpp, ExLlamaV3 and ComfyUI, with a scheduled job queue. It's the small end of the range I build across, from models on my own GPUs to frontier APIs.
04How I work
The tools live inside Claude and ChatGPT, on the laptop or phone people already carry. Our senior superintendent used them on site before any field training.
Google sign-in and a role check sit in front of every tool. That is what made it safe to connect real business systems.
Staff tell me what worked and what didn't, and requests become working tools in 1–2 days. Adoption kept growing after launch because the people using the tools shaped them.
Years of implementing and migrating finance, HR, fleet and payment systems, including several QuickBooks Desktop to Online moves, meant the AI layer sat on connected data.
05Case study
Two pages on the problem, what I built, how I rolled it out from the office to the field, and what changed. Includes the adoption numbers, the in-house shop drawings and the earned-revenue report.
Download the case study