Birju Kadakia — CEO of Rec Technologies (recreation tech; early Uber employee who helped launch Uber Eats worldwide, then led product at The Athletic) — interviewed by Angelina on TwoSetAI (56 min). Rec is the AI layer being slid under 25 years of incumbent recreation software: registration, court reservations, memberships, and refunds for municipal parks & rec departments, with 100+ cities live.

The activity economy is a $400B market running on pen and paper

  • “Local activity” — youth sports, swim lessons, summer camps, leagues, gyms, senior-center programs — is a $400B/year US economy; 50% of the country’s recreational spaces are operated by municipalities
  • Three-sided: space owners (cities, YMCAs, sports complexes), operators (coaches, volunteers, lifeguards, league runners), and consumers — all coordinated on paper
  • The consumer touchpoint is the printed summer-camp catalog that arrives in the mail; parents flip through it, strategize, and race to book at 8am on signup day

Three problems hiding in one catalog

  • Cities spend weeks hand-assembling activity catalogs that then get printed
  • Nobody wants to flip through a paper catalog (younger parents may not even check the mail)
  • Booking day is stressful and opaque — will the system hold up, is the right section picked, is the plan ready?

Why the wedge starts B2B, with the recreation professionals

  • Consumers suffer once a season; operators and space owners are bogged down year-round (recruiting staff, marketing programs, taking attendance on paper, data-entry, financial reporting for funding)
  • Rec sells to the people running recreation; the consumer experience improves as an outcome
  • “It doesn’t need to be branded as AI… it will just get really great and you’ll be happier with your city because you can be more active”

The consumer-facing agent vision

  • An agent knows what the kids did last summer, their ages and interests — it assesses the catalog, proposes an itinerary, and offers to book it on booking day
  • FastTrack (in the product now) lets families plan bookings in advance — users tweet “I love FastTrack” when a city goes live
  • Step one of the whole company was getting the activity calendar into a structured, AI-accessible format: “putting that calendar somewhere accessible” so agents — theirs or others’ — can act on it

Live in 100+ cities — “the Shopify for the city’s recreation resources”

  • SF parks & rec runs tennis/pickleball court reservations on Rec; Dolores Park courts carry “powered by Rec” QR signs like OpenTable placards
  • Indianapolis goes live in the fall — 800,000 residents each getting a Rec profile (“everyone sees Indy parks and rec as the city’s best-kept secret, and we don’t want to be a secret anymore”)
  • Rec ID lives in the wallet — a “recreation passport”: preferences, waivers, medical info, emergency contacts, allergies, digitized check-in
  • His analogy for the back office: a hospital’s EHR, or Toast for restaurants — “we’re providing all the software that runs the recreation ecosystem”

The refund agent: codify the policy instead of inferring it

  • Rain cancels swim classes, tennis reservations, and camps — every class has its own cancellation/refund policy, too many edge cases for hand-built automation
  • Their pattern: an LLM reads each city’s policies ONCE and writes deterministic code that codifies them — the LLM is used at build time, not per-call, so inference errors are designed out of the hot path
  • Operators see the policy in plain language and an audit log of every section the agent would refund — “make it feel very trustworthy for a city to use AI”

Skills → workflows: the agent architecture

  • Refunding a user is one skill; deciding whether it’s rainy enough is another; a workflow chains them on triggers (“weather API says rain → cancel this set + auto-refund”)
  • Today they sharpen individual skills; a dedicated team is building the end-to-end workflow builder that ties them any way a city wants
  • The moat is knowing how recreation actually operates — “we train the system to do things the way a recreation operator would”

The vertical wedge vs. raw Claude

  • The host’s framing: a city admin could pipe its policies into Claude — Rec’s answer is that data alone isn’t enough; you need access (schedules, customers, payments) plus industry taste for the true decision boundary
  • Harvey and Lorra (legal AI) build the same thing: workflows plus an understanding of how the domain works
  • “OpenAI is not going to knock door to door to each city and ask ‘can I help you take care of your recreation process?’” — the horizontal labs may become recreation assistants one day, but Rec is also the operating system of the data and interactions that make AI useful there

Is vertical SaaS dead? “Some software is dead”

  • Basic website builders and undifferentiated tools die — no reason to buy them
  • Survivors own: (1) the valuable private data (schedule, customer preference, compliance/health), (2) payments and trust — high-stakes, high-trust transactions like a $3,000 summer camp, and (3) infrastructure for equity and access (personalized pricing, financial-aid early access)
  • A really enterprising city could vibe-code its own version — “but it would be more expensive than working with Rec”

First data: solve a micro-problem first

  • 2022: pickleball was blowing up and every court needed better software — Rec solved that one problem, went live with 10 cities, and suddenly owned schedule + consumer data
  • “We earned the right to grow from there” — his tip to founders: pick a small problem before you solve every problem

Hallucination is a layered problem

  • Principle: “you can’t trust inference at the moment of inference” — build trust as a layer around it
  • Layer 1 — human-in-the-loop onboarding: refund suggestions first (“here are the last 10 cancellations, here’s what I’d do and why”), the admin turns on automation once they trust it — modeled on how you’d onboard an employee
  • Layer 2 — don’t infer where you don’t have to: policy codified as deterministic code avoids hallucination altogether
  • Layer 3 — the messy middle (customer comms, finding a picnic table): eval conversations with Langfuse, plus people manually scoring batches of 50 — “everything is changing so fast, some of it is just scrappy”
  • Real accuracy numbers by task: deterministic paths 100%, refund suggestions 60–80% at tuning time, recommending a two-kid schedule with different preferences ~40–50% — the bar is being “meaningfully more useful than a world without AI,” like a great travel agent’s first draft

First ten customers: the unglamorous grind

  • “You cold call a hundred, you talk to five, and you get in” — their actual first customer found THEM (the market pulled the product), the next five came from references
  • ~3 months to 10 customers; pushing 100 the following year
  • On rejection: never ignore the customer who says it sucks, but most rejection is the wrong customer at the wrong time — take the signal and adjust the product, the pitch, or the target

A 35-person team in the AI age

  • “We can do more with less” — true across the board; everyone is more productive
  • Roles are colliding: PMs ship code every day, the best engineers manage many agents and have become mini PMs, designers need PM-grade communication and structured thinking
  • The core of each function survives: engineers own architecture, scaling, and queueing; designers stay pixel-perfect; PMs keep judgment and prioritization over hundreds of customer asks
  • Hiring changed: interview prompts assume AI use (“use AI to X”), engineering loops now include a full product interview — “deep at one thing, reasonably deep at the others”

What engineers and data scientists should learn

  • The job of an engineer today is the job of the great engineering manager you once had: technical advice, unblocking, and communication — but aimed at a team of agents
  • Don’t measure by token usage — “it feels like a false proxy; we’re going to look and laugh at ourselves one day” — measure shipping velocity
  • The failure mode to avoid: generating 100K lines of code that bogs down code review and “nothing gets out” — everyone was actually less efficient

Founder advice: tune the company to speed

  • “The single word is speed — time is your most precious commodity… I wish I learned that faster” is the takeaway from every look back, right or wrong
  • Culture mechanics: if you learn something, share it right away; if you want to change direction, talk about it tonight, not next week
  • He uses Claude daily as an EA — routines that brief him for meetings, summarizing what to act on; he still decides what’s impactful, then makes each thing easier
  • Book rec: Amp It Up by Frank Slootman (Snowflake/ServiceNow) on setting goals and building a fail-fast culture — plus read for enjoyment, not just self-help

“Where vertical AI companies will truly succeed are when the founders and the teams are deeply embedded in that ecosystem — because it’s going to be really hard for someone like Claude or OpenAI to train on how the work happens and build the workflows around it.”