←
andreas cleve457 posts
andreas cleve avatar

andreas cleve

@andreascleve

@corti_ai. the only ai lab built for healthcare.

◷ Joined November 2010▧ 25 Photos and videos
Sandhya avatar
⟳ andreas cleve retweeted
Sandhya@sandhya·Apr 11

the New Software strategy checklist: how many of these is your company already executing on? 🤖 *full* capability parity* between what humans and agents can do 🚀 CLI/SDK/APIs for agents to provision and configure your product *without HITL* 🎓 evolving *battle-tested skills* that encode your best practitioners' judgment 💰benchmarks that show you are *high performance, low cost* compared to general purpose agents (thanks to vertical models/custom databases) 🪨 stable user/agent interfaces and harnesses that *outlast model changes* For CEOs: less than 3/5 by the end of Apr 2026 and your co is ngmi

Sandhya @sandhya Apr 10 Article The New Software: CLI, Skills & Vertical Models In the era of agent experience, performance will become the new competitive advantage for great SaaS companies In January 2025, we first started debating the “death” of software. Anthropic had just

andreas cleve avatar
andreas cleve@andreascleve·Apr 2

we’ve trained Symphony to work the way a trained human coder works step by step, codebook open, citing the exact sentence that justified every decision so clinicians can verify it, auditors can check it, and no patient gets harmed by a guess. It’s the agentic way of doing it.

andreas cleve avatar
andreas cleve@andreascleve·Apr 2

every major AI lab took a shot at this treating it as bolt-on problem. But by vertically specializing, we beat all of them. 📄Read the Paper: arxiv.org/abs/2603.29709 It’s been tested on real hospital data across 5 datasets, 2 countries, inpatient and outpatient peer reviewed.

andreas cleve avatar
andreas cleve@andreascleve·Apr 2

but it’s not just about billing - coding governs a lot of things that are downstream in healthcare like national health registries. in one national dataset bad coding meant 3 in 4 suicide attempts went uncounted according to the patient documentation. Coding is hard.

andreas cleve avatar
andreas cleve@andreascleve·Apr 2

medical coding is the unsexy infrastructure layer running all of healthcare every billing decision, every insurance claimC every patient record is built on 72,000 codes assigned “by hand”, and even the best humans get it wrong which costs $36B/year in the US alone

andreas cleve avatar
andreas cleve@andreascleve·Apr 2

Medical Coding has a new leader we just beat anthropic, google, openai, amazon and oracle by 25% or more. Symphony for Medical Coding is live 🧵

Post attachment
andreas cleve avatar
andreas cleve@andreascleve·15 Oct 2025

Available now for: 🎙️ Dictation (/transcribe) 💬 Ambient (/stream) Rated Premier, covering every major specialty from cardiology to psychiatry. We also added: 🗣️ Commands — hands-free voice control for templates, navigation, and automation ✍️ Spoken punctuation & formatting — so transcripts read like true medical notes

andreas cleve avatar
andreas cleve@andreascleve·15 Oct 2025

Corti’s new French ASR model was trained and validated to overcome these challenges: • +36% improvement in Medical Term Recall • 2.11% Word Error Rate (WER) • <1% Character Error Rate (CER) (Benchmarks normalized; punctuation and case sensitivity ignored - so please test on your own data)

andreas cleve avatar
andreas cleve@andreascleve·15 Oct 2025

🔐 Real-world data is limited. Healthcare audio is strictly regulated under CNIL and GDPR, which means less clinical speech data for training. That scarcity makes domain-specific modeling essential.

andreas cleve avatar
andreas cleve@andreascleve·15 Oct 2025

📚 Medical French is complex and Latin-heavy. Terms like œdème, myocarde, and laryngectomie have morphologies and pronunciations that diverge from their written forms. This confuses models that are primarily trained on everyday speech.

andreas cleve avatar
andreas cleve@andreascleve·15 Oct 2025

🧠 Medical speech mixes languages. Clinicians constantly code-switch between French and English: “Le patient a eu un check-up et un CT scan.” Generic ASR models degrade quickly when bilingual patterns appear mid-sentence.

andreas cleve avatar
andreas cleve@andreascleve·15 Oct 2025

💬 Accents and regional variation are vast. French healthcare spans speakers from France, Canada, and across the francophone world - each with unique rhythm and pronunciation patterns. Robust ASR needs to understand all of them equally well.

andreas cleve avatar
andreas cleve@andreascleve·15 Oct 2025

🎧 French blends words together. Through liaison and elision, word boundaries disappear: “les antécédents médicaux” sounds like lezantécédan médicaux. (Sorry for my terrible French) That makes it difficult for ASR models to segment correctly, especially in busy hospital environments.