Muzzammil Khan

Switching from prompt libraries to agent frameworks? Be ready for real learning. We’ve been deep in the weeds with Cloud Code SDK, learning agent patterns, sub-agents, orchestration flows (parallel, sequential, hybrid), and every “promising” pattern shared. It all looks amazing in a POC… until you try to use it in a real solution. The “perfect” orchestration pattern you hoped for still needs tuning, retries, and guardrails. There’s no silver bullet. Just steady experimentation, many failures, and slow, consistent learning about what works for your codebase. Don't build a fancy framework with routers, memory/state management, etc., until you can develop features with sub-agents & an orchestrator (and some patterns). After weeks of trial and refinement, we’re finally seeing consistent results (still ways to go). If you’re on this path, embrace the journey. Progress is real, but it takes time. One more: Experiment widely during R&D between Claude, Codex, and Antigravity (Google), but for production systems, pick one ecosystem and stay anchored. #AI #Engineering #Agents #CloudCode #LLM

21

Yes, the EHR Will Kill your AI Startup. Epic launched a native AI scribe. Athenahealth has four ambient vendors embedded natively. Oracle is building AI-first and has their "Clinical Agent". Athenahealth recently announced 2 way text and voice AI... $1.6B was invested in ambient AI startups in 2025. A lot of that money is not coming back. Similar to the spooking of SaaS companies that is happening as a result of Claude, there's no question in my mind that the post-AI VC splurge contraction is coming. As generative AI foundational models get better and better, context is key. And context for agentic action comes from data. And that data is viciously guarded throughout healthcare in large part by the system of record...so startups will die in waiting while the "roadmaps" eerily similar to the most successful use cases start to roll out natively in the leviathan EHRs of the world. But beyond the culling of the features masquerating as products, I’m more interested to see where the survivors go next (or already going?) Care orchestration. Post-acute transitions. Patient engagement between visits. Deeply clinical workflows tied to new payment models or reimbursement requirements. Entirely new models of AI first clinician backed care. Work that the EHR was never built to serve. The startups that survive won't be competing with Epic and Athena and Oracle. They'll be building something the system of record never imagined, and something our healthcare system desperately needs. I will say, no matter how you slice the landscape, this isn't just a market dynamic. It's a regulatory failure that innovation is so contingent on closed systems. Read more on Techy Surgeon ↓

587

118 Comments

Such an exciting year ahead. Healthcare AI is finally crossing from digitization to cognition shifting value from administrative drag to clinical and scientific productivity, with revenue cycle intelligence as the first scalable proof of ROI. Today, administrative functions consume roughly one-quarter of total U.S. healthcare spending making AI-driven cognitive automation one of the few levers capable of reallocating dollars back to care, not just cutting costs. With average hospital operating margins near 2%, even a 30–50 basis-point improvement from AI-driven clinical intelligence can translate into a 15–25% relative margin uplift without adding beds, clinicians, or capital spend. #MBA #Healthcare #AI

6

Both Epic and Oracle Health recently announced incorporation of AI into current and future EHRs. 𝘊𝘰𝘰𝘭 𝘚𝘵𝘶𝘧𝘧 𝘕𝘰𝘸: 𝘈𝘐 𝘊𝘩𝘢𝘳𝘵𝘪𝘯𝘨 (https://lnkd.in/g9dRe68x), a video by Bradley Fox, MD, Director of Clinical Informatics, announced a number of current and planned future AI enhancements to Epic's widely used hospital EHR that he says will "simplify tasks across the health system from the back office to the point of care". I found these interesting: Clinical summaries include care that took place elsewhere. Patient notes will "translate medical jargon into approachable language". Future patient facing capabilities will explain test results, explain billing balances, answer questions and help make appointments. It extracts findings and recommends follow-ups from radiology reports. He says the ambient dictation "is smart enough to skip details that aren't relevant". In the future, it will "cue up orders and diagnoses based on the conversations". It helps nurses "with their flowsheet documentation" and "end of shift notes". Finally, for billing, it suggests codes and can draft appeal letters when claims are denied. Oracle Health just introduced its new Health EHR 𝘧𝘰𝘳 𝘱𝘩𝘺𝘴𝘪𝘤𝘪𝘢𝘯𝘴 in which, it says, "AI is seamlessly embedded into every layer —from data and services to the user interface—to create a scalable, integrated solution for personalized, data-driven care." I found these interesting features reading the Oracle product brief (https://lnkd.in/gAPTunZx) which is clear that it includes future features. It is integrated with health information exchanges, including Oracle Health Seamless Exchange, Oracle Health Data Intelligence, and TEFCA Qualified Health Information Networks (QHIN) "to retrieve and consolidate longitudinal patient records". The list of current QHINs is at https://lnkd.in/g85B4CBb. It includes Epic Nexus and Oracle Health has been "accepted into the project planning and testing phase". Physicians can interact with the EHR's embedded Clinical AI Agent 𝘣𝘺 𝘷𝘰𝘪𝘤𝘦 "across multiple functions, including chart review, documentation, and task automation". From the day's schedule it "provides a 'fish-eye view' AI-generated summary of each patient's visit purpose and relevant background, with highlighted sources". I've written about Retrieval Augmented Generation or RAG which Oracle apparently plans to use (although the brochure is aimed at physicians and doesn't say this) "for the clinician to be able to use voice commands to pull near real time, patient-specific insights from a multi-source catalogue, seamlessly integrating clinical intelligence into the workflow." Sources will include drug databases, clinical guidelines, pharmacy records, and immunization histories. I've long advocated offering physicians a timeline view of their charts, and Oracle says its new EHR will provide one.

34

2 Comments

A rewarding part of my job at Kaiser Permanente is working with our customers to help them understand both the opportunities and the risks of AI technologies to promote the health of our members. I thought it would be useful to share some of our key insights with the broader community. 1) Healthcare in the US is struggling to keep up with rising demand, leading to increased costs of care, provider burnout and access delays. 2) AI presents an opportunity to alleviate some of these consequences by driving operational efficiency, supporting our clinical workforce and enhancing pt access. 3) But AI risks must also be attended to - including the risk of bias, inaccurate outputs, public mistrust, and an increasingly complex regulatory landscape. 4) Health care organizations must emphasize and demonstrate the safe and responsible use of AI in their organizations. 5) But the risks of AI must also be weighed against the risk of inaction and the status quo. Read the full article below: https://lnkd.in/gFq3QQwh

153

12 Comments

Delivery models are rarely replaced; they are layered. ​Just as Scrum wrapped around Waterfall cores, we are now seeing a new layer emerge as AI moves from the periphery to the center of product strategy. ​When AI starts influencing outcomes, "delivery speed" loses its status as the primary metric. The new challenge is behavioral integrity. ​Scrum defines the build. N8–GEMS defines the life of the system thereafter. ​N8 shifts the mindset to AI as a living entity. ​GEMS provides the governance and observability required for trust. ​We’re no longer just building software; we’re stewarding systems that learn. It’s complex, high-stakes, and happening mid-flight. ​Which of these feels closest to your personal voice?

33

1 Comment