we build ai apps, iphone apps and agent systems in weeks. two of them are ours.
larry labs is an ai product studio in new york. we ship fast because the studio runs on a software factory: agents that build, test, review and monitor the work, with a human deciding and reviewing. the apps we build are ai-native: the model does the back-office work and never blocks the user. here is the proof.
wydtn: a city's nightly events paper, printed by agents
- problem
- new york publishes thousands of events a day across hundreds of venue sites, calendars and newsletters. no person can read them all, so every listings app is partial and stale.
- what we built
- an ios and web app fed by a chain of agents: a scout finds sources, a crawler reads them at 4am, an extractor turns pages into events, a second model reviews the extractor, code gates drop adult, corporate and out-of-area listings, a ranker retrains nightly, and a printer publishes at 11:59pm and proves the edition reached the site. ai runs only in the back office; a reader never waits on a model.
- outcome
- about 3,000 events in every edition's seven-day window, from a record of 32,000 (september 2026). the model that checks work is never the one that produced it.
- time to ship
- first prompt to the app store in 27 days (2026-07-31 to 2026-08-27).
larry match: an ai matchmaker with a character
- problem
- dating apps make people swipe. we wanted one where an ai matchmaker gets to know you in conversation and finds your matches, and where that matchmaker can be trusted around real people.
- what we built
- a web app with an ai character (larry), a reasoning layer behind him, and a matching pipeline. before opening it up we wrote 380 adversarial scenarios (manipulation, minors, a person in crisis, scams, harassment) and ran them against the live reasoning layer.
- outcome
- the first run found 19 unsafe responses. we fixed the 103 hardest scenarios first (6 failures, then 1, then 0), then re-ran all 380 and got 0 (june 2026).
- time to ship
- the safety sweep, fixes and validation ran inside one week.
- live
- larrymatch.com (web beta).
the operating system: agents that run a company
- problem
- one person cannot build, ship, monitor and support two consumer products. we needed the company itself to run on agents, with a human deciding and reviewing rather than typing.
- what we built
- a set of role agents on our own repos: a fixer that drains the bug board, a conductor that runs a spec end to end unattended, an observability layer where every process declares what "working" means and a sentinel checks it every five minutes, and a spend ledger that reads every vendor's billing daily. routine work goes to a smaller model; a rule sends it there.
- outcome
- over four weeks ending 2026-09-28: 91% of operations completed with no human intervention, 37 changesets per human-hour, one human. routing routine work to the smaller model cut the ai bill 12% in the week measured, with a blind judge confirming quality held (17 of 21).
- time to ship
- built alongside the two products, march to september 2026.
- live
- nothing public; we demo it on a call.
how we start
every engagement starts with a two-week discovery sprint, $10,000. we learn your workflow, build a working prototype on the riskiest assumption, and hand you a scope and a fixed quote for the build. the sprint fee is credited in full against a build signed within 60 days.
what a build includes
- your repo and your cloud from day one
- an eval harness and cost telemetry
- human approval gates where the model touches a person
- handoff docs
- 30 days of support
- a monthly partner tier if you want us to stay.
– johnny · j@larrylabs.co