Start from the real problem
People describe problems in their own words. HTO Assistant takes a plain message from staff, checks for an existing ticket, then adds to it or files one. Nobody has to learn the tracker first.
Forward Deployed Engineer · Applied AI
Ben Downey
I put software into real operations and keep it running.
For 25 years I have run the production system, the customers and the support for a sports league SaaS. Since April 2026 I have also built and run a multi-agent AI system every day, with routing, review gates, tests and spend controls.
Animated diagram. Ben hands a request to a secretary, who spawns tasks into a live queue of up to ten. A hub session sends each task to one to three agent lanes running the model that fits the job: Opus for review, Sonnet for code and the browser, Haiku for lookups. Every result passes a tests and review gate before it ships, and a failed check goes back to the hub for a fix.
01 · The agent system
A multi-agent system I built and run every day.
One hub session decides and routes. Cheaper models do the bulk work, chosen per job. Every agent has a written lane, a list of what it never handles, and the credentials it may read.
145+tasks completed per month
$14k+API-equivalent AI usage per month
25B+tokens processed per month
18,000+automated tests
- Hooks block pushes to production repos, reads of secrets, and spend on metered APIs.
- A commit gate refuses code without a green test run.
- A cross-model review registry decides which reviewer may see which class of data: public, internal, personal, source code, customer data.
- A written spend policy, plus an audit command that checks for paid API use.
- A single-user web cockpit in TypeScript to approve, reject, pause or stop agents.
- So far: 3,000+ commits since April 2026, 2,000+ Python files, 10+ agent definitions each with its own lane, 70+ skills, and 60+ hook scripts that block unsafe moves.
02 · 25 years in production
Co-founder of HomeTeamsOnline. Running production since 2000.
Sports league websites and management software. The company has served tens of thousands of sports organizations across the US and Canada. I lead engineering and product, and I keep it running.
25yearsin production since 2000
1.9M+registrations processed
3M+games scheduled
25M+page views per month
- Lead engineer on the main web application: the production platform behind the company's league websites and management software, built by a small team.
- Built the Node.js and Express API that runs next to the main web application, including a read-only public API used by league customers.
- Built HTO Assistant, a Discord bot where staff report problems in plain language. It finds the existing JIRA ticket, adds to it or files a new one, with one Claude call per message.
- I query the production SQL Server database directly, with bounded read-only queries.
03 · Working with customers
Next to the customer, owning the result.
For 25 years I have run the production system, the customers and the support for the same product. That shapes how I work.
Read the live data, carefully
Answers come from the production database, read with bounded, read-only queries. Not from guesses.
Plans that survive review
My plan to take the whole product to a mobile app on one shared API went through a second, code-grounded review. It found 17 issues. All 17 were fixed in the plan, starting with the security order for legacy write endpoints.
Own it after launch
Shipping is the middle, not the end. I have kept the same production system running for 25 years, and I run the agent system with tests and spend controls every day.
04 · Writing
Explaining AI to engineers, and to everyone else.
I also run DXdev, where I write about building with AI agents. Two tracks: one for people who build software, one for people who just want the work to run.
For engineersHow the system is put together and what keeps it safe, written for the people who would build one.
Open the trackopens in a new tab
For everyone elsePlain-language stories about work that gets stuck, and how practical AI makes it easier to run.
Open the trackopens in a new tab
Three to start with
Field noteStaging Had Twelve Players Per Group and Production Had 47A clean deploy, then two data-shape bugs in the first hours. Why watching production is part of done.opens in a new tab
Everyday AIThe Schedule Bug That Pointed Back to March 4, 1973A date number turned into text, and text sorts differently. One small change put a schedule out of order.opens in a new tab
Field note136 Bounced Emails Joined Back to Real Customer OrdersOne week of receipts from running a legacy platform with AI agents, misses included.opens in a new tab
05 · Experience
The short version.
Tools I use
- Python
- JavaScript
- Node.js
- Express
- TypeScript
- SQL Server
- Astro
- Claude Code
- JIRA
- agent design
- evals of agent behavior
- cost control
- production support
- incident recovery
Languages
English and Portuguese (fluent)
- 2026 now
Builder, multi-agent AI system
Hub and lanes, hooks, review gates, tests and spend controls. Python, plus a TypeScript cockpit.
- 2026
HTO Assistant
Discord bot that turns plain staff messages into JIRA tickets. Python, one Claude call per message.
- 2016
HTO public API
Node.js and Express API next to the main web application, with a read-only public API for league customers.
- 2003
B.S. Computer Science
Minor in Electrical Engineering, Magna Cum Laude. Rensselaer Polytechnic Institute.
- 2000 now
Co-Founder, Engineering and Product, HomeTeamsOnline
Production system, customers and support for a sports league SaaS, for 25 years and counting.
06 · Contact
Let's talk.
Send me a note here, or find me on LinkedIn.