Carey Phelps

This month, 101 GenAI features a happy customer, Shalini Gupta, discussing Asima Health's utilization of agentic AI in healthcare. In this engaging conversation, Shalini Gupta shares her unconventional journey from academia to entrepreneurship, detailing her motivations for founding Asima Health, a company focused on cancer diagnostics. She discusses the technical challenges faced in developing innovative diagnostic tools, the importance of measuring success through clinical milestones, and her vision for democratizing access to cancer screening. Shalini also highlights the role of AI in enhancing diagnostic capabilities, her personal reflections on failure and success, and the balance between work and personal life. The conversation concludes with her recommendations for reading and influential figures in her life. Also, a big thank you to Shefali Lathwal, Smita Agrawal, Prashanth Aditya Susarla & Praveen Dua for their amazing work with Asima's team. Watch the full video here: https://lnkd.in/gZ3geZcu How are you utilizing Agentic AI to enhance healthcare? Please comment.

89

1 Comment

In today’s noisy landscape, where everything is labeled “AI,” it has become essential to distinguish between AI as a buzzword and AI as a true operating layer for business. As highlighted in a16z’s Big Ideas 2026, the next era is not about isolated copilots or scattered AI assistants. It is about deeply integrated, agentic AI infrastructure—systems that are coherent, reliable, and embedded across workflows, decision-making, and execution. The age of fragmented experimentation is ending. The new paradigm favors vertical, domain-aware AI that solves complex business problems end-to-end, quietly and efficiently, without adding cognitive or operational noise. In this environment, precision is no longer optional. It becomes the core prerequisite for scale, trust, and sustained performance. Organizations that treat AI as infrastructure—not a feature—will define the next competitive frontier. For thoughtful, tailor-made insights on this shift, follow Viacheslav Kalaushin.

5

2 Comments

From AI 1.0 to AI 3.0: Over thirty years in, I’ve seen AI cycle through multiple reinventions. Today’s multimodal LLMs look new and gigantic—but their mechanics extend ideas refined decades ago. Five enduring paradigms remain technically fundamental: 1. Feature Extraction and Representation — Compression and basis decomposition. Pre-deep-learning pipelines used methods such as autoencoders, convolutions, FFTs, PCA, and clustering to project high-dimensional signals into compact, structured manifolds. They defined invariances, sparsity, and locality—concepts that modern encoders, tokenizers, and latent-space embeddings still rely on. 2. Genetic Algorithms (GAs) — Gradient-free structural search. Evolutionary operators performed discrete and continuous optimization across non-convex spaces. Current parallels include neuroevolution, population-based training, and evolutionary prompt tuning for LLM architectures. 3. Bayesian Belief Networks (BBNs) — Probabilistic inference and causal dependency modeling. Nodes captured conditional structure; posterior updates implemented sequential belief revision. Their logic reappears in probabilistic programming, Bayesian deep learning, and causal transformers. 4. Local Learning — Piecewise estimation and neighborhood adaptation. RBFNs, locally weighted regression, and Gaussian-process kernels realized localized generalization. Their kernels mirrored biological on-center/off-surround receptive fields—"sombrero-hat" Gaussian profiles emphasizing proximal samples while inhibiting distant ones. Modern analogs: mixture-of-experts routing, meta-learning, and attention locality. 5. Fuzzy Logic — Continuous membership functions and differentiable rule blending. Fuzzy sets implemented graded truth; today’s softmax layers, attention weights, and mixture gating reproduce the same mathematics at scale. These principles are not obsolete—they are the mathematical foundations that modern architectures rediscover at scale. #ArtificialIntelligence #MachineLearning #LLM #DeepLearning #FoundationalModels

7

Why Small Language Models (SLMs) Are Powering High-Precision Prior Authorization in Healthcare Prior authorization remains one of the most expensive and friction-heavy workflows in healthcare. The challenge isn’t lack of data — it’s precision. Large, general-purpose language models are impressive, but prior authorization demands something different: deep domain understanding, deterministic behavior, and consistent outcomes across payer-specific rules. This is where Small Language Models (SLMs) are proving to be far more effective. SLMs trained specifically on prior authorization workflows, payer policies, CPT and ICD mappings, medical necessity rules, and historical authorization outcomes can operate with significantly higher precision. They reduce hallucination risk, handle edge cases more reliably, and align better with compliance and audit requirements. More importantly, SLMs enable intent-driven automation. They can identify authorization requirements early, extract the exact clinical evidence needed, assemble payer-specific submissions, and flag gaps before a request ever reaches the payer. The result is fewer delays, fewer resubmissions, and faster patient access to care. Another advantage is deployability. SLMs are lighter, faster, and easier to operate in private or hybrid environments — a critical requirement for healthcare organizations that care about data control, latency, and regulatory boundaries. At Contiinex, we see prior authorization evolving from a manual, reactive process into a predictive, AI-assisted workflow powered by specialized SLMs — not generic models. #PriorAuthorization #HealthcareAI #SLMs #RevenueCycleManagement #DigitalHealth #Contiinex

11

How are the public for-profit's thinking about tech and AI adoption? Deploying at scale - Ardent Health. AI scribe at 85% adoption, 35% documentation time reduction. Medical wearables reducing mortality by 15% and LOS by ~0.3 days. Virtual care expansion to 2,000+ rooms. The single-instance Epic platform is the enabler and I can remember this being called out as a strategic advantage to the org in its S-1. Most ambitious strategic vision: HCA Healthcare. Three-domain framework (administrative, operational, clinical). The "holy grail" of clinical AI using HCA's proprietary database for pattern recognition to support physician decision-making. EHR transition to standardize datasets. $400M resiliency program leveraging digital transformation. Most practical near-term deployment: UHS. Hippocratic AI partnership for agentic AI in post-discharge care. Revenue cycle AI for claims appeals. Behavioral health intake/referral AI. Patient safety technology for behavioral settings. Filton acknowledged difficulty in precise quantification but cited reduced headcount and improved readmissions. Most transformative (maybe?) infrastructure play: Community Health Systems. Oracle ERP generating ~$50M in savings in Year 1. Single item master across the enterprise. Quarterly AI functionality updates from Oracle cloud. AI in claims appeals, autonomous coding, prior authorization. Also deploying AI-augmented virtual patient sitters and ambient listening. Most "structural" framing: Tenet Healthcare. Technology-enabled expense management through global business centers and clinical throughput optimization. Emphasized this is "modernization" not traditional cost-cutting. Separately, digital integration with payers for electronic data exchange and administrative simplification is driving working capital improvements. My question is - when do these efforts meaningfully move margin profiles. I do think it's sooner rather than later.

37

1 Comment

ViVE Session: Patient-Centered AI - Creating Tech People Trust Panelists: Randall Rutta (National Health Council) Sneha Dave (Generation Patient) Michelle Monaco (Wheel) Chethan Sarabu, MD, MD (Cornell Tech) This was an interesting session on how patients should be part of AI development. Patients do not speak the same language as physicians, yet many models are evaluated only by clinicians. The patient experience is much bigger than a diagnosis - a sick working mom is not just a “case.” Context matters and AI tools can play a larger role by understanding context. Interesting moment: ~60% of attendees said they already use AI tools to navigate health - from wellness to chronic care. AI is expanding beyond hospital walls. Important concerns raised: - Informed consent cannot be a checkbox. It needs deeper conversations. - What happens when patients bring AI scribes or tools into visits? This is coming and providers should be ready with this. - Transparency, safety, and data protection must be clear to both patients and physicians. - LLMs may underrepresent underserved populations. Synthetic data may hide rare conditions. - Mental health AI tools need stronger evidence standards. *Big takeaway*: Trust is not built by capability alone. It is built by transparency, safety, and real patient partnership in how AI is designed and deployed. #Vive2026

17

2 Comments

HeyDoc AI has unveiled India’s first 𝗩𝗼𝗶𝗰𝗲-𝗔𝗴𝗲𝗻𝘁-𝗥𝘂𝗻 𝗛𝗼𝘀𝗽𝗶𝘁𝗮𝗹 𝗖𝗵𝗮𝗶𝗻 in partnership with Clearmedi Healthcare (A Morgan Stanley Company), setting a new benchmark in Agentic AI driven healthcare. Powered by 𝗪𝗲𝗹𝗹𝗻𝗲𝘀𝘀𝗚𝗣𝗧 𝗩𝗼𝗶𝗰𝗲 𝗔𝗴𝗲𝗻𝘁𝘀 including 𝗦𝘆𝗺𝗽𝘁𝗼𝗺 𝗧𝗿𝗶𝗮𝗴𝗲 𝗔𝗴𝗲𝗻𝘁𝘀, 𝗔𝗽𝗽𝗼𝗶𝗻𝘁𝗺𝗲𝗻𝘁 𝗦𝗰𝗵𝗲𝗱𝘂𝗹𝗶𝗻𝗴 𝗔𝗴𝗲𝗻𝘁𝘀, 𝗣𝗮𝘁𝗶𝗲𝗻𝘁 𝗙𝗼𝗹𝗹𝗼𝘄-𝗨𝗽 𝗔𝗴𝗲𝗻𝘁𝘀, 𝗖𝗮𝗿𝗲 𝗣𝗹𝗮𝗻 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁 𝗔𝗴𝗲𝗻𝘁𝘀, 𝗛𝗼𝘀𝗽𝗶𝘁𝗮𝗹 𝗢𝗽𝗲𝗿𝗮𝘁𝗶𝗼𝗻𝘀 𝗔𝗴𝗲𝗻𝘁𝘀, and 𝗜𝗻𝘀𝘂𝗿𝗮𝗻𝗰𝗲 𝗖𝗼𝘃𝗲𝗿𝗮𝗴𝗲 𝗔𝘀𝘀𝗲𝘀𝘀𝗺𝗲𝗻𝘁 𝗔𝗴𝗲𝗻𝘁𝘀 this deployment enables true end-to-end, agentic healthcare delivery across hospital networks. But this is just the beginning. Tomorrow at 𝗗𝗛𝗡 𝗖𝗶𝘁𝘆 𝗠𝗲𝗲𝘁𝘂𝗽 𝗗𝗲𝗹𝗵𝗶, we are also unveiling the 𝗪𝗲𝗹𝗹𝗻𝗲𝘀𝘀𝗚𝗣𝗧 𝗖𝗮𝗿𝗲 𝗔𝗴𝗲𝗻𝘁 𝗦𝘁𝘂𝗱𝗶𝗼 enabling 𝗵𝗼𝘀𝗽𝗶𝘁𝗮𝗹 𝗰𝗵𝗮𝗶𝗻𝘀 𝗮𝗻𝗱 𝗱𝗶𝗴𝗶𝘁𝗮𝗹 𝗵𝗲𝗮𝗹𝘁𝗵 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 𝘁𝗼 𝗱𝗲𝘀𝗶𝗴𝗻, 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗲 𝗮𝗻𝗱 𝗱𝗲𝗽𝗹𝗼𝘆 𝘁𝗵𝗲𝗶𝗿 𝗼𝘄𝗻 𝗔𝗜 𝗮𝗻𝗱 𝘃𝗼𝗶𝗰𝗲 𝗮𝗴𝗲𝗻𝘁𝘀 𝗮𝗹𝗶𝗴𝗻𝗲𝗱 𝘄𝗶𝘁𝗵 𝘁𝗵𝗲𝗶𝗿 𝗰𝗹𝗶𝗻𝗶𝗰𝗮𝗹 𝗦𝗢𝗣𝘀. Built on our temporal agentic memory layer and multi-agent orchestration engine, the Studio transforms Hospital Chains and Digital Health Companies from AI adopters into AI operators. Over 𝟭𝟬𝟬 𝗖𝗫𝗢𝘀 𝗳𝗿𝗼𝗺 𝗹𝗲𝗮𝗱𝗶𝗻𝗴 𝗵𝗼𝘀𝗽𝗶𝘁𝗮𝗹 𝗰𝗵𝗮𝗶𝗻𝘀 𝗮𝗻𝗱 𝗱𝗶𝗴𝗶𝘁𝗮𝗹 𝗵𝗲𝗮𝗹𝘁𝗵 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 will be present as we move from conversation to category creation. #AgenticAI #VoiceAI #DigitalHealth #HealthcareInnovation #WellnessGPT #DHN

22

I spent the last 5 days at the India AI Summit 2026 I will skip some of the positives for now. We learn more from friction. Here’s where the math didn’t add up for me: 1. Irresponsible AI deployment is being underestimated. Are current AI executions increasing productivity? Unknown. Is productivity already being promised and committed in contracts? Absolutely. 2. Embeddings ≠ understanding. I saw too many startups putting embeddings on top of raw data and calling it intelligence. Very little discussion on business context, workflow integration, or decision risk. 3. Hallucinations are real. Are the current executions hallucinating? Yes. Are hallucinations systematically measured and addressed ? No. In agriculture. In healthcare. In sensitive sectors. 4. “We trained our own LLM.” For startups to claim this casually, without acknowledging the required data scale and compute power, signals narrative inflation. 5. Deterministic workflows and probabilistic systems. Witnessed startups plugging probabilistic LLM outputs into deterministic business processes compliance, financial ops, medical flows without redesigning the business workflows around them. I’m starting to see early symptoms of a bubble forming not in the foundational AI research, but in the execution layer of AI in India. #IndiaAIImpactSummit2026 #ResponsibleAI #AIOrchestration

8