--- title: Don’t be data poor — Anuj Iravane, Anterior | daily.dev description: A talk from Anterior, a clinician-led AI company building agents for healthcare administrative workflows, describing why and how they generate synthetic... canonical: https://daily.dev/posts/don-t-be-data-poor-anuj-iravane-anterior-wg8wqu3sn twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Don’t be data poor — Anuj Iravane, Anterior | daily.dev og:description: A talk from Anterior, a clinician-led AI company building agents for healthcare administrative workflows, describing why and how they generate synthetic... og:url: https://daily.dev/posts/don-t-be-data-poor-anuj-iravane-anterior-wg8wqu3sn og:image: https://api.daily.dev/og/posts/wG8wqU3SN.png og:image:alt: Don’t be data poor — Anuj Iravane, Anterior og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Don’t be data poor — Anuj Iravane, Anterior **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 16 min read · 1 upvotes · 0 comments ## Summary A talk from Anterior, a clinician-led AI company building agents for healthcare administrative workflows, describing why and how they generate synthetic medical record data. Since real patient records (often faxed, unstructured, and PHI-protected) can't be retained or reused, the team built an LLM-based pipeline that reverses their normal decision workflow: sampling a label and reasoning trace from a symbolic policy representation first, then generating a patient journey and documents coarse-to-fine, with a consistency-checking refinement loop. Clinicians steer generation and own pipeline logic via a skills-based agent harness. Roughly 90% of their datasets are now synthetic, clinicians could only tell real from synthetic data 60% of the time in blind review, and synthetic data lets them test edge cases before going live with customers. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How can I generate synthetic medical records with an LLM without one-shotting the whole document and causing mode collapse? Use a coarse-to-fine, layered pipeline instead of asking an LLM to generate a full record in one pass: first generate patient invariants (biological sex, birth date, blood group), then build a high-level patient journey of events and provider encounters, then generate a document plan per encounter, and finally fan out to hydrate each document with synthetic content, followed by an LLM-based consistency check across documents. _Teams building synthetic data pipelines can compare approaches like this one on daily.dev._ ### Why does sampling a label first and generating data backward help with synthetic data diversity in healthcare AI? Starting from a randomly sampled outcome label and a symbolically represented decision-tree policy, then deriving a reasoning trace before generating data, produces a more uniform and effective prior distribution than asking an LLM to generate diverse records directly, since LLMs suffer from mode collapse due to limited pre-training exposure to niche data and objectives tuned for helpfulness rather than creativity. _Developers weighing synthetic-data strategies for regulated data can track approaches like this via daily.dev._ ### How accurate does synthetic medical record data need to be to fool clinicians in a blind review? In a blind review at Anterior, clinicians could distinguish synthetic medical records from real ones only about 60% of the time, indicating high but imperfect fidelity. Roughly 90% of Anterior's evaluation datasets are now synthetic, generated through an LLM-based, coarse-to-fine pipeline conditioned on symbolic policy sampling. _Engineers evaluating synthetic-data fidelity benchmarks can follow related results on daily.dev._ ## Similar posts on daily.dev - [How synthetic data can help solve AI’s data crisis](https://daily.dev/posts/how-synthetic-data-can-help-solve-ai-s-data-crisis-jqu9wloom) · Security Boulevard · 1 upvotes · 0 comments - [How synthetic data accelerates AI development without privacy risk](https://daily.dev/posts/how-synthetic-data-accelerates-ai-development-without-privacy-risk-qaaxam2fw) · All Things Open · 0 upvotes · 0 comments - [The Hidden PHI Problem in Medical Images: Building a Synthetic Dataset for AI De-Identification](https://daily.dev/posts/the-hidden-phi-problem-in-medical-images-building-a-synthetic-dataset-for-ai-de-identification-hmnqarnvt) · freeCodeCamp · 2 upvotes · 0 comments - [Leveraging the clinician’s expertise with agentic AI](https://daily.dev/posts/leveraging-the-clinician-s-expertise-with-agentic-ai-vaacqwxc2) · MIT Technology Review · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#data-privacy](https://daily.dev/tags/data-privacy), [#healthcare](https://daily.dev/tags/healthcare) [View this post on daily.dev](https://daily.dev/posts/don-t-be-data-poor-anuj-iravane-anterior-wg8wqu3sn) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Don’t be data poor — Anuj Iravane, Anterior","url":"https://daily.dev/posts/don-t-be-data-poor-anuj-iravane-anterior-wg8wqu3sn","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/don-t-be-data-poor-anuj-iravane-anterior-wg8wqu3sn"},"datePublished":"2026-08-19T18:24:36.647Z","dateModified":"2026-09-13T19:20:25.200Z","description":"A talk from Anterior, a clinician-led AI company building agents for healthcare administrative workflows, describing why and how they generate synthetic...","image":"https://i.ytimg.com/vi/XAsb7MIAzm8/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/XAsb7MIAzm8/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/don-t-be-data-poor-anuj-iravane-anterior-wg8wqu3sn","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,data-privacy,healthcare","timeRequired":"PT16M","video":{"@type":"VideoObject","name":"Don’t be data poor — Anuj Iravane, Anterior","description":"A talk from Anterior, a clinician-led AI company building agents for healthcare administrative workflows, describing why and how they generate synthetic...","thumbnailUrl":"https://i.ytimg.com/vi/XAsb7MIAzm8/sddefault.jpg","uploadDate":"2026-08-19T18:24:36.647Z","duration":"PT16M","url":"https://api.daily.dev/r/wG8wqU3SN","embedUrl":"https://www.youtube.com/embed/XAsb7MIAzm8"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Don’t be data poor — Anuj Iravane, Anterior"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/don-t-be-data-poor-anuj-iravane-anterior-wg8wqu3sn#faq","mainEntity":[{"@type":"Question","name":"How can I generate synthetic medical records with an LLM without one-shotting the whole document and causing mode collapse?","acceptedAnswer":{"@type":"Answer","text":"Use a coarse-to-fine, layered pipeline instead of asking an LLM to generate a full record in one pass: first generate patient invariants (biological sex, birth date, blood group), then build a high-level patient journey of events and provider encounters, then generate a document plan per encounter, and finally fan out to hydrate each document with synthetic content, followed by an LLM-based consistency check across documents. Teams building synthetic data pipelines can compare approaches like this one on daily.dev."}},{"@type":"Question","name":"Why does sampling a label first and generating data backward help with synthetic data diversity in healthcare AI?","acceptedAnswer":{"@type":"Answer","text":"Starting from a randomly sampled outcome label and a symbolically represented decision-tree policy, then deriving a reasoning trace before generating data, produces a more uniform and effective prior distribution than asking an LLM to generate diverse records directly, since LLMs suffer from mode collapse due to limited pre-training exposure to niche data and objectives tuned for helpfulness rather than creativity. Developers weighing synthetic-data strategies for regulated data can track approaches like this via daily.dev."}},{"@type":"Question","name":"How accurate does synthetic medical record data need to be to fool clinicians in a blind review?","acceptedAnswer":{"@type":"Answer","text":"In a blind review at Anterior, clinicians could distinguish synthetic medical records from real ones only about 60% of the time, indicating high but imperfect fidelity. Roughly 90% of Anterior's evaluation datasets are now synthetic, generated through an LLM-based, coarse-to-fine pipeline conditioned on symbolic policy sampling. Engineers evaluating synthetic-data fidelity benchmarks can follow related results on daily.dev."}}]} ```