{
  "video": {
    "id": "XAsb7MIAzm8",
    "title": "Don’t be data poor — Anuj Iravane, Anterior",
    "duration": 1006,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "XAsb7MIAzm8",
    "title": "Don’t be data poor — Anuj Iravane, Anterior",
    "one_liner": "Anterior generates its own synthetic medical records by running its inference workflow backwards — sampling a label, then a reasoning trace from a symbolic decision-tree policy, then building documents coarse-to-fine — because PHI contracts forbid keeping the real data it most needs for evals.",
    "summary": "Anuj Iravane leads AI at Anterior, which builds agents for healthcare admin workflows (prior authorization, payment integrity, HEDIS) that amount to policy-guided decision-making over highly unstructured data — mostly scanned fax bundles, since ~70% of medical communication still happens by fax. The data is PHI: it can't be retained, reused, or even derived from, so no eval dataset survives, and in healthcare 95% accuracy isn't good enough. Their answer is to generate the data themselves, but not by one-shotting a 300-page record with an LLM, which mode-collapses; instead they reverse the forward task — sample a random label, deterministically sample a reasoning trace from a policy modelled as a symbolic decision tree, then build a record layer by layer (patient invariants → patient journey → per-encounter document plan → fan-out document generation → eval-driven refinement). Roughly 90% of their datasets are now synthetic, clinicians distinguish synthetic from real only ~60% of the time in blind review, and clinicians own the pipeline directly because every stage is a skill file on an internal agent harness.",
    "key_points": [
      "Anterior's workflows (prior auth, payment integrity, HEDIS) all reduce to 'policy-guided decision-making over highly unstructured data' — scanned fax bundles with bad handwriting, tables, checkboxes, key-value pairs and images, often 300+ pages, modelling an entire clinical trajectory; ~70% of medical communication still goes by fax.",
      "The constraint that drives everything: PHI can't be retained, reused, or derived from, and most contracts also prohibit redacting/anonymizing and keeping derivative copies — so nothing persists as a dataset, while healthcare baselines mean 95% accuracy is not good enough.",
      "One-shotting synthetic records with an LLM fails: it's like asking for a novel in one shot, and LLMs mode-collapse on diversity because this data barely appears in the pre-training corpus and pre/post-training objectives reward helpfulness, not creativity or diversity.",
      "The core technique is reversing the forward task: instead of data + policy → reasoning trace → label, sample a random label, then a reasoning trace, then generate the data backwards from that conditioning input.",
      "Anterior models policies explicitly as decision trees / symbolic representations (example: a CPAP medical-necessity policy). That improves accuracy and consistency in LLM execution, and lets them deterministically sample reasoning traces per outcome — a far more uniform prior distribution than sampling from an LLM, and one that covers rare edge cases a 200-case customer sample never would.",
      "Generation is coarse-to-fine and layered: patient invariants (biological sex, birth date, blood group) → an ordered list of events and providers called the 'patient journey' → a document plan per provider encounter → parallel fan-out to hydrate the actual documents. This keeps input and output prompt payloads token-efficient and scales to long journeys without blowing context windows.",
      "A refinement loop applies evals as feedback, including an LLM consistency check across documents to catch contradictions introduced by the parallel fan-out; because generation started from labels, a round-trip check confirms the record matches the task inputs/outputs, so labels are correct by construction and expensive ground-truth labelling is skipped.",
      "Everything stays in plain text/markdown — no rendering to PDF, because state-of-the-art PDF parsers already turn complex PDFs into clean markdown, so generation and evaluation both happen in the text domain.",
      "Domain experts own the pipeline two ways: human-in-the-loop steering at every generation step (clinicians take interesting production cases and steer toward look-alikes), and the whole pipeline modelled as skills on an internal generic agent harness — a clinician adds a new document type for a new customer by writing a new skill file, with no engineering changes.",
      "Results: ~90% of Anterior's datasets are already synthetic; in blind review clinicians could only distinguish synthetic from real about 60% of the time; datasets are now created just-in-time for customer deployments instead of waiting on customer data, so edge cases are simulated and tested before go-live."
    ],
    "takeaways": [
      "Reverse your inference workflow to generate data: sample the outcome first, then the reasoning trace, then generate the input — you get diversity and correct labels by construction.",
      "Sample diversity from a distribution appropriate to your use case rather than asking an LLM for it; a symbolic representation of your policy or decision logic gives you a uniform, deterministically samplable prior.",
      "Emulate the real-world process that produced your data (for medical records, generation during provider encounters), and build coarse-to-fine in layers so prompts stay token-efficient and long records don't overflow context.",
      "Don't one-shot long documents with an LLM, and don't bother rendering to PDF — modern parsers put you back in markdown anyway, so generate and evaluate in the text domain.",
      "Give domain experts the keys: expose the pipeline as skills plus human-in-the-loop steering so clinicians (not AI engineers) own the logic and drive recursive self-improvement.",
      "Apply this beyond PHI — anywhere data is ephemeral, sensitive, or expensive to label, generate it yourself instead of waiting for customer data."
    ],
    "topics": [
      "synthetic data",
      "evals",
      "healthcare ai",
      "privacy",
      "agents",
      "skills",
      "unstructured data",
      "human-in-the-loop"
    ],
    "tools": [
      "Anterior",
      "Sequoia",
      "NEA",
      "LLMs",
      "Cynthia",
      "PDF parsers",
      "Claude-style skills / internal agent harness"
    ],
    "quotes": [
      {
        "text": "So, so what this talk is about is like what do you do when the dataset you most need is also the data you're least allowed to keep.",
        "at": "02:54",
        "url": "https://www.youtube.com/watch?v=XAsb7MIAzm8&t=174s"
      },
      {
        "text": "in healthcare the the baselines for accuracy are just exceptionally high. 95% is not good enough.",
        "at": "02:05",
        "url": "https://www.youtube.com/watch?v=XAsb7MIAzm8&t=125s"
      },
      {
        "text": "it's like imagining if you wouldn't ask an LLM to write a novel for you in one shot, right? So, it's the same reason why you wouldn't use an LLM to just one shot a synthetic record for you.",
        "at": "04:14",
        "url": "https://www.youtube.com/watch?v=XAsb7MIAzm8&t=254s"
      },
      {
        "text": "I feel like skills are really an amazing interface between AI engineers and domain experts, especially in vertical AI.",
        "at": "13:48",
        "url": "https://www.youtube.com/watch?v=XAsb7MIAzm8&t=828s"
      },
      {
        "text": "In a blind review, clinicians were not able were only able to distinguish synthetic from real about 60% of the time.",
        "at": "14:36",
        "url": "https://www.youtube.com/watch?v=XAsb7MIAzm8&t=876s"
      }
    ],
    "words": 3533
  },
  "summary_url": "/#XAsb7MIAzm8",
  "transcript": {
    "html": "/transcripts/XAsb7MIAzm8.html",
    "txt": "/transcripts/XAsb7MIAzm8.txt",
    "vtt": "/transcripts/XAsb7MIAzm8.vtt"
  }
}