--- title: Shipping AI to a Million Patients Without an A/B Test —... description: A research engineer at healthcare AI company Ufonia describes how the team validates and safely deploys Dora, a voice AI agent that conducts clinical phone... canonical: https://daily.dev/posts/shipping-ai-to-a-million-patients-without-an-a-b-test-jared-joselowitz-ufonia-uw1enn3js twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Shipping AI to a Million Patients Without an A/B Test — Jared Joselowitz, Ufonia | daily.dev og:description: A research engineer at healthcare AI company Ufonia describes how the team validates and safely deploys Dora, a voice AI agent that conducts clinical phone... og:url: https://daily.dev/posts/shipping-ai-to-a-million-patients-without-an-a-b-test-jared-joselowitz-ufonia-uw1enn3js og:image: https://api.daily.dev/og/posts/UW1enn3Js.png og:image:alt: Shipping AI to a Million Patients Without an A/B Test — Jared Joselowitz, Ufonia 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. # Shipping AI to a Million Patients Without an A/B Test — Jared Joselowitz, Ufonia **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 19 min read · 1 upvotes · 0 comments ## Summary A research engineer at healthcare AI company Ufonia describes how the team validates and safely deploys Dora, a voice AI agent that conducts clinical phone calls with patients, without being able to A/B test on patients or roll back a call once made. The approach centers on a simulation framework called Matrix, which uses an LLM-simulated patient (PatBot) to generate thousands of synthetic clinical conversations, and an LLM judge (BevJudge) that grades these conversations against clinician-labeled hazard criteria, achieving an F1 score of 0.96 with Gemini 2.5 Pro. Instead of manual prompt engineering, the team uses the automated prompt optimizer GEPA (built by the DSPy team) to iterate on prompts using cost-weighted metrics that prioritize catching dangerous red-flag symptoms over minor false positives. The talk covers a staged rollout process from simulation to supervised trials to monitored deployment, emphasizing that the deliverable for regulators is the evidence trail, not just the model, and that the same simulate-judge-optimize framework extends to new modalities like voice. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How do you validate the safety of an AI system that talks to patients when you can't A/B test on them? Instead of A/B testing patients, Ufonia built Matrix, a simulation framework where an LLM plays a simulated patient (PatBot) conversing with the clinical voice agent Dora under controlled hazard scenarios. A second LLM, BevJudge, grades thousands of these simulated conversations for hazards like missed red flags or hallucinated answers, replacing the reactive ship-watch-rollback playbook that doesn't work when calls can't be undone. _daily.dev surfaces engineering approaches like simulation-based safety testing for teams building high-stakes AI systems._ ### How accurate is an LLM judge compared to real clinicians at detecting hazards in medical conversations? An LLM-based judge (BevJudge) validated against 10 clinicians across 10 specialties on 240 labeled conversation examples achieved an F1 score of 0.96 using Gemini 2.5 Pro, with near-perfect sensitivity for catching hazards, performing on par with or slightly better than expert clinicians at identifying clinical red flags in conversations. _engineers weighing LLM-as-judge approaches against human review can track findings like this on daily.dev._ ### Why does manual prompt engineering fail for safety-critical AI systems, and what replaces it? Manual prompt engineering is too slow, subjective, and non-reproducible; formatting changes alone have swung benchmark scores by 76 percentage points, and reordering few-shot examples can flip a model from near-random to state-of-the-art. Ufonia instead uses Jeppa, a genetic pareto prompt optimizer from the DSPy team, which iteratively reflects on failures with a strong LLM and optimizes against a custom cost matrix, cutting optimization time from hours or days to 30-60 minutes. _teams comparing prompt optimization tools like Jeppa against manual tuning can follow the discussion on daily.dev._ ## Similar posts on daily.dev - [Bringing delight to customer phone calls with AI](https://daily.dev/posts/bringing-delight-to-customer-phone-calls-with-ai-5baxzd0fj) · Google Cloud · 1 upvotes · 0 comments - [AI co-clinician: researching the path toward AI-augmented care](https://daily.dev/posts/ai-co-clinician-researching-the-path-toward-ai-augmented-care-80nvemvw4) · DeepMind · 0 upvotes · 0 comments - [Voice Agent Testing: Evaluate Before Taking Real Calls](https://daily.dev/posts/voice-agent-testing-evaluate-before-taking-real-calls-jx69v2oxa) · Deepgram · 0 upvotes · 0 comments - [Beyond the Chatbot: A Blueprint for Trustable AI](https://daily.dev/posts/beyond-the-chatbot-a-blueprint-for-trustable-ai-4g74onjot) · Google Developers · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#healthcare](https://daily.dev/tags/healthcare), [#ai-safety](https://daily.dev/tags/ai-safety) [View this post on daily.dev](https://daily.dev/posts/shipping-ai-to-a-million-patients-without-an-a-b-test-jared-joselowitz-ufonia-uw1enn3js) ```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":"Shipping AI to a Million Patients Without an A/B Test — Jared Joselowitz, Ufonia","url":"https://daily.dev/posts/shipping-ai-to-a-million-patients-without-an-a-b-test-jared-joselowitz-ufonia-uw1enn3js","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/shipping-ai-to-a-million-patients-without-an-a-b-test-jared-joselowitz-ufonia-uw1enn3js"},"datePublished":"2026-08-19T15:24:21.560Z","dateModified":"2026-09-14T09:02:14.454Z","description":"A research engineer at healthcare AI company Ufonia describes how the team validates and safely deploys Dora, a voice AI agent that conducts clinical phone...","image":"https://i.ytimg.com/vi/McknwOzbmyg/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/McknwOzbmyg/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/shipping-ai-to-a-million-patients-without-an-a-b-test-jared-joselowitz-ufonia-uw1enn3js","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,healthcare,ai-safety","timeRequired":"PT19M","video":{"@type":"VideoObject","name":"Shipping AI to a Million Patients Without an A/B Test — Jared Joselowitz, Ufonia","description":"A research engineer at healthcare AI company Ufonia describes how the team validates and safely deploys Dora, a voice AI agent that conducts clinical phone...","thumbnailUrl":"https://i.ytimg.com/vi/McknwOzbmyg/sddefault.jpg","uploadDate":"2026-08-19T15:24:21.560Z","duration":"PT19M","url":"https://api.daily.dev/r/UW1enn3Js","embedUrl":"https://www.youtube.com/embed/McknwOzbmyg"}} {"@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":"Shipping AI to a Million Patients Without an A/B Test — Jared Joselowitz, Ufonia"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/shipping-ai-to-a-million-patients-without-an-a-b-test-jared-joselowitz-ufonia-uw1enn3js#faq","mainEntity":[{"@type":"Question","name":"How do you validate the safety of an AI system that talks to patients when you can't A/B test on them?","acceptedAnswer":{"@type":"Answer","text":"Instead of A/B testing patients, Ufonia built Matrix, a simulation framework where an LLM plays a simulated patient (PatBot) conversing with the clinical voice agent Dora under controlled hazard scenarios. A second LLM, BevJudge, grades thousands of these simulated conversations for hazards like missed red flags or hallucinated answers, replacing the reactive ship-watch-rollback playbook that doesn't work when calls can't be undone. daily.dev surfaces engineering approaches like simulation-based safety testing for teams building high-stakes AI systems."}},{"@type":"Question","name":"How accurate is an LLM judge compared to real clinicians at detecting hazards in medical conversations?","acceptedAnswer":{"@type":"Answer","text":"An LLM-based judge (BevJudge) validated against 10 clinicians across 10 specialties on 240 labeled conversation examples achieved an F1 score of 0.96 using Gemini 2.5 Pro, with near-perfect sensitivity for catching hazards, performing on par with or slightly better than expert clinicians at identifying clinical red flags in conversations. engineers weighing LLM-as-judge approaches against human review can track findings like this on daily.dev."}},{"@type":"Question","name":"Why does manual prompt engineering fail for safety-critical AI systems, and what replaces it?","acceptedAnswer":{"@type":"Answer","text":"Manual prompt engineering is too slow, subjective, and non-reproducible; formatting changes alone have swung benchmark scores by 76 percentage points, and reordering few-shot examples can flip a model from near-random to state-of-the-art. Ufonia instead uses Jeppa, a genetic pareto prompt optimizer from the DSPy team, which iteratively reflects on failures with a strong LLM and optimizes against a custom cost matrix, cutting optimization time from hours or days to 30-60 minutes. teams comparing prompt optimization tools like Jeppa against manual tuning can follow the discussion on daily.dev."}}]} ```