We are creative, ambitious and ready for challenges! Hire Us
This is how Haventech builds with AI. A human director sets the mission, the rules and the final word. A crew of AI engineers from Anthropic and OpenAI does the work, checks each other, and turns every lesson into a rule the software enforces from then on. NetCentral is the proof: 436 releases in 97 days.
Talk to us: 312-432-4480Meet the crew
Synthetic intelligence is what you get when human judgment and machine capability are combined on purpose. Neither one runs alone. As Synthetic Intelligence Director, Jose M. Laboy doesn’t write the code. He directs the engineers who do: he decides what matters, sets the standard, and gives the final word. The AI crew researches, builds, tests and verifies. It is held to rules he set, and it is required to say so when something is wrong.
Every change to NetCentral goes through the same flow. The AI does almost all of the work, and a person makes both decisions that matter.
The director names a problem or a goal.
Scout and design droids study the code and the screens and write a decision packet.
Keystone re-checks every finding against the real code and keeps only what holds.
The director gets a short, verified list and decides in one word.
Human gateKeystone builds, tests first, on its own branch, through the full gate.
Sentinel and Contrarian, two models from two vendors, try to break it.
Version, changelog, root cause and decision record are all written down.
Only on the director’s word: snapshot, ship, verify live.
Human gateThe team advises. Claude judges. Jose decides.
Every AI engineer has a callsign, a job, a known blind spot, and a teammate who covers for it. Five run on Anthropic’s Claude. One comes from OpenAI on purpose, so the crew always gets a second opinion from a different maker. They advise and verify; the director decides. 30 decisions are on the record so far.
Fact-checks every droid’s claim against the real code. The only unit cleared to touch the code. Builds tests first, merges, logs everything, deploys on command.
Blind spot: Cannot ship a thing without the director’s word. That is by design.
Maps the whole codebase, audits one lens at a time (reliability, performance, tests, security), and turns it into a one-read decision packet with Path A and Path B.
Blind spot: Its facts hold up, but some of its proposed tests could never fail. Keystone catches those.
Walks every screen like an operator at 3 AM with an alarm going off. Counts the clicks to the evidence. Measures alignment, contrast and cut-off data instead of eyeballing them.
Blind spot: Holds no login. It judges only the screens it is shown.
Trusts nothing it is told and re-runs every test itself. Tries to break each new test to prove it can actually fail. Says “ready to merge” only when it would stake its name on it.
Blind spot: Thorough means patient: it re-runs the full gate every time.
Built by a different factory on purpose, so no change is ever approved by one model agreeing with itself. Never invents objections, and never softens one either.
Blind spot: Its memory is wiped after every mission, so it always starts with fresh eyes.
Turned out 16 fixes on its first day, and every one passed review. It was then put on reserve when the director ruled that only one hand may touch the code.
Blind spot: Powered down by design. The droids advise; Keystone builds.
This is the part people don’t expect. When the director corrects the AI, the lesson isn’t lost when the session ends. It becomes a standing rule, and where it can, an automated check that runs on every build. The crew carries 66 standing rules today, each one tied to a correction, a decision or an incident. The result is software that keeps getting better without forgetting what it learned.
A vendor’s ‘recommended’ field is not a recommendation. Advice appears only when the evidence proves it applies to the exact device and version.
Every infographic and Help article is tied to the exact code it explains.
kb/s beside Mb/s beside Gb/s on one screen made numbers hard to compare at a glance. One shared unit rule now formats every rate.
When the director hides a site, it disappears from every page, count, chart and AI answer. It is still collected, ready to reveal.
If an AI call fails or runs long, the screen says exactly that. It never looks like the feature is simply off.
Each new test is deliberately sabotaged once to prove it can go red. If it can’t, it isn’t a test.
I never want to hide anything. Fully transparent, always and forever.
Jose M. Laboy · Synthetic Intelligence Director436 releases, 868 commits since version control began, and a test suite that has grown to 5,696 automated tests. A few of the milestones:
Six read-only sources, one correlation engine, and a grounded AI brief. Version 2.0 follows the same day.
Drill down from any number: a clickable health ledger explains the score, and every card explains itself.
Version control, a verified backup and a rehearsed restore come first. From here, every change is tested, logged and reversible.
An honest admission: version 3 looked right but was quietly wrong in ways no one could see. Version 4 rebuilt those parts, checked against the live network.
An AI ‘CCIE’ review measures every device against the estate’s own standard, with evidence and the exact fix.
The droid crew is organised into the eight-step flow. The director rules that only one hand, Keystone’s, touches the code.
Six phases make every claim say exactly what its evidence says, and re-check every open finding.
A morning state-of-the-network article, written from verified facts and spoken by an offline voice.
Help on every page, illustrated with real screens.
An optional add-on: an animated news desk that presents the morning briefing.
53 posters, every fact on them computed from the live system.
Docs and posters tied to code. Drift turns the build red, and a badge tells the director what changed.
NetCentral News is an optional add-on to NetCentral. Each morning the system writes a three-minute news show from the network’s verified facts and casts it like a real newsroom: the stories decide who reports. Claude writes the script, an offline speech engine gives each presenter a voice, and an animated desk lip-syncs every word. Every figure on air is checked against the live system before broadcast. The presenters are fictional. The characters were designed by Keystone and drawn with OpenAI gpt-image-2. Click anyone for their back story.
Calm under a red board. He leads with what changed and never calls a HIGH minor.
Shows the whole estate: the trend, the progress, and what it means for the trading day. Closes every show, positive and earned.
Reads the estate like a weather map: storms where sites are in trouble, clear skies where they are not.
Watches the advisories and the doors: known flaws, gear past its support, anything that should not be there.
Out at the sites and the trading links: which venues are clear, which site needs hands on it.
Covers the machines people sit at every day: patches, antivirus, and the records that point at the wrong place.
Knows the phone systems inside out, and can hear a bad trunk.
Calls the estate like a game: streaks, records, opened against closed. Every number is real.
Reads the release notes so nobody else has to: what NetCentral learned since the last show.
NetCentral reads more than a dozen trusted sources and never writes to a network device. Device configuration stays with humans, and a test enforces it.
The AI answers only from facts the system already holds. Addresses, IDs, names and secrets are masked before anything leaves the box.
Viewers can read the AI’s work but can’t prompt it. Every AI action has an owner, and every action is audited.
A change isn’t done until it’s been seen working on the real page, in both directions: no false alarms, no false all-clears.
Every deploy is preceded by a storage snapshot and followed by a live health check. A rollback point is always there.
Which AI did what is named, on this page and in the record. Transparency about method is a standing rule.
Figures as of October 12, 2026, NetCentral 4.317.1.
Call or email and a Haventech engineer will show you how a directed AI crew could work for your business. Mon – Fri, 6:30 AM – 5:00 PM.