{
  "video": {
    "id": "WJRdLNhrsLQ",
    "title": "How to build an AI-Native Health Company — Dan Feng, Maven Clinic",
    "duration": 1039,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "WJRdLNhrsLQ",
    "title": "How to build an AI-Native Health Company — Dan Feng, Maven Clinic",
    "one_liner": "Maven Clinic's Dan Feng lays out how a digital health company went AI-native in two years — not by buying tools, but by changing hiring, planning horizons, code review and reliability engineering around the fact that implementation is now cheap and judgement is expensive.",
    "summary": "Feng argues there's no playbook for 'AI native' and defines it as three things: use AI internally wherever a task would otherwise be done manually or delegated, build AI into the product to improve UX and cut operational cost, and — most importantly — change culture, process and ways of working to maximise what AI offers. He walks through the consequences at Maven: senior engineers stop delegating implementation, planning collapses to 2–4 week sprints because 3–6 month plans can't survive unknown model releases, and code review is redesigned for engineers now writing thousands of lines a day. He closes on reliability, arguing hallucination can't be eliminated affordably so you triage which failures are acceptable, cross-check high-stakes flows with multiple models, and run integration tests many times demanding a ~90% pass rate.",
    "key_points": [
      "Maven Clinic started its AI journey ~2 years ago and built 'Maven Intelligence', an orchestration layer across all products enabling AI for everyone in the company and for clients.",
      "Adoption is segmented into three groups: early adopters (just enable tools and get them to share), the majority (build shared AI infrastructure, easy-to-use tools, listen to feedback), and slow adopters (understand their concerns, but be crystal clear where the company is heading).",
      "Meet engineers where they are on tooling — most of Maven used Cursor last year, many switched to Claude Code this year, and the company supports both.",
      "The delegation model broke: senior engineers who've figured out the solution now implement it with AI instantly rather than hand it to another engineer, because delegation means more overhead and less efficiency; new hires must be able to solve problems independently.",
      "Hiring shifted to people genuinely interested in AI, engineers who understand the product (PM/engineer boundaries are blurring), and those who handle deep systems understanding and ambiguous problems — 'where AI lands off'. Performance reviews now explicitly ask what you've done on the AI side.",
      "Planning changed shape: one-year thinking is inspirational only (assume models can do anything by then), the real focus is what ships in 2–4 weeks, and 3–6 month mid-term goals are deliberately de-emphasised because nobody knows what models will do by then. PRDs/TDDs are now one or two pages as communication artifacts to iterate on, not pages and pages.",
      "AI coding tools were adopted lowest-risk-first — unit tests and documentation, which are easy to verify — to build confidence and construct their own rules, skills and guardrails, before mandating them across engineering; now AI does essentially all implementation and engineers focus on reviewing, architecting and evaluation.",
      "Code review adapted to volume (hundreds of lines/day → thousands): engineers can self-identify that a PR needs no review and merge it while remaining accountable; reviewed PRs should be under 500 lines; stacked PRs let big features be split; 'rubber stamp' blind approval is the worst case because it gives false confidence.",
      "Reliability is triaged by failure cost: a 1-in-1,000 failure on appointment scheduling is tolerable (the user re-clicks), but reimbursement claims are not — a $200 claim paid as $50 escalates immediately — so receipts are reviewed by different models and only proceed when the models agree, otherwise the user is offered a human agent.",
      "Release process: hundreds of integration tests covering all known use cases, each run many times (passing once isn't good enough with an LLM) demanding a consistently high pass rate such as 90%; after launch an auto-eval system scores every conversation against predefined rubrics, plus a dedicated human group spot-checks conversations and reviews ~20% when new features launch.",
      "The stated goal not yet reached: fully automating the software lifecycle end-to-end, including AI monitoring live traffic to catch issues early and fix them automatically."
    ],
    "takeaways": [
      "Start AI coding adoption on verifiable, low-risk work (unit tests, docs) to build confidence and accumulate your own rules and guardrails before mandating it everywhere — and when people opt out, treat that as a signal to learn why.",
      "Shrink the planning horizon: make the one-year vision directional only, commit concretely to the next 2–4 weeks, and stop writing long PRDs/TDDs since being wrong two weeks later is cheap and switching gears is fine.",
      "Redesign code review for AI-scale output rather than pretending the old process scales: cap reviewed PRs at ~500 lines, stack PRs for big features, let engineers self-certify simple PRs while staying accountable, and actively hunt out rubber-stamping.",
      "Classify your AI failures by cost before shipping — accept cheap-to-retry failures, and for irreversible ones (money, claims) run multiple different models and only proceed on agreement, with a human handoff as the acceptable fallback.",
      "Change what you test and what you reward: run every integration test many times against a pass-rate threshold (~90%) instead of once, and add 'what have you done with AI?' to performance reviews so multiplied impact is rewarded."
    ],
    "topics": [
      "ai-native",
      "engineering-culture",
      "ai-coding-tools",
      "code-review",
      "evals",
      "reliability",
      "hiring",
      "healthtech"
    ],
    "tools": [
      "Maven Clinic",
      "Maven Intelligence",
      "Cursor",
      "Claude Code",
      "Jira"
    ],
    "quotes": [
      {
        "text": "Like a tractors aren't to replace farmers, but the farmers who can operate the tractor will replace the ones who cannot.",
        "at": "01:30",
        "url": "https://www.youtube.com/watch?v=WJRdLNhrsLQ&t=90s"
      },
      {
        "text": "With AI, building is super fast. It's probably couple minutes you can get it done.",
        "at": "07:36",
        "url": "https://www.youtube.com/watch?v=WJRdLNhrsLQ&t=456s"
      },
      {
        "text": "One thing we really want to avoid is a rubber stamp, we call it. Means like people submit code review, you cannot really do anything to it. You just say blindly approve it. This is the worst case, we should really avoid because that's just give us false confidence.",
        "at": "12:30",
        "url": "https://www.youtube.com/watch?v=WJRdLNhrsLQ&t=750s"
      },
      {
        "text": "The really awkward part is mid-term goals. Those like a three months, six months. It's very hard to plan these days. The reason is I don't know what AI models will be capable in three months.",
        "at": "09:01",
        "url": "https://www.youtube.com/watch?v=WJRdLNhrsLQ&t=541s"
      }
    ],
    "words": 3134
  },
  "summary_url": "/#WJRdLNhrsLQ",
  "transcript": {
    "html": "/transcripts/WJRdLNhrsLQ.html",
    "txt": "/transcripts/WJRdLNhrsLQ.txt",
    "vtt": "/transcripts/WJRdLNhrsLQ.vtt"
  }
}