Brandon T.

Carlos Guzman’s #OpenMed walkthrough is a fantastic proof point: bring a healthcare-tuned model to your data, run it at scale, and only pay while it’s running. 1M clinical patient notes, 4 GPU’s, 5m30s, ~$2–3. This is how we vault healthcare forward with AI —by unlocking real use cases at low cost. #Snowflake #CortexAI

12

Unleashing AI for Clinical Data: GPT-4o Powers 3,458 Patient Records into i2b2 Key Takeaways: • 3 disparate data models (three synthetic AnVIL datasets) automated into a unified pipeline [Ensures broad applicability across spreadsheet and bespoke formats] • 3,458 fake patient records ingested seamlessly into i2b2 via ChatGPT-4o–generated Python scripts [Delivers immediate cohort discoverability for researchers] • Two-stage transform: data dictionaries → intermediate common format → i2b2 ontology [Guarantees semantic consistency using NCBO’s standard terminologies] • Original clinical files converted into i2b2 facts with no manual coding tweaks [Minimizes human error and accelerates deployment] • Prompt reuse slashes incremental setup for each new dataset [Drives scalable ETL workflows and substantial time savings] #ClinicalData #ETL #LLM Link: https://lnkd.in/edEvSzid

1

Enjoying this year's Databricks Data & AI Summit. Lakebase is the most promising architecture I've seen yet for separating compute and storage for OLTP databases. The proof will be in the execution and adoption, but as with many Databricks initiatives, I'm excited to see this in action.

51

1 Comment

Worth a read: Moving healthcare from siloed, batch data to real-time, multimodal, AI-ready architectures—so AI can reason and act, not just run point solutions. It nicely highlights how scaling AI depends on modernizing the data backbone: unified pipelines, strong governance, and constructs like knowledge graphs that make context truly usable. Energized to help accelerate this journey at Mayo Clinic. #HealthcareAI #DataArchitecture #DataGovernance https://lnkd.in/gfdmXtEK

48

1 Comment

One of the most impactful design decisions I made on a healthcare project was implementing Medallion Architecture (Bronze → Silver → Gold) in Databricks. At CVS Health, we were ingesting large volumes of pharmacy claims, eligibility, and provider data coming from PBMs like OptumRx in multiple formats (JSON, XML, Parquet). What I actually did: 🔹 Bronze Layer I landed raw claims and eligibility feeds exactly as received from S3, without transformation. This helped us with auditability, reprocessing, and HIPAA traceability when downstream issues occurred. 🔹 Silver Layer Here I applied real business logic — Deduplication of claims Eligibility validation Standardization of member, provider, and plan identifiers PHI masking rules This layer became the trusted operational dataset used across teams. 🔹 Gold Layer I built analytics-ready tables for reporting, SLA monitoring, and downstream dashboards. These were optimized for Redshift/Snowflake consumption and Power BI reporting, with performance tuning and partitioning. Why this worked in production: Reduced reprocessing effort when source data changed Clear separation of concerns between raw, clean, and business data Made debugging claims issues much faster Helped us stay compliant with HIPAA and GDPR This architecture wasn’t theoretical — it directly supported near real-time eligibility checks and helped reduce claim adjudication errors by ~30%. Sometimes good architecture isn’t about fancy tools — it’s about designing data layers that teams can actually trust. #DataEngineering #Databricks #DeltaLake #MedallionArchitecture #HealthcareData #AWS #Snowflake

3

Ambient scribe isn’t a moat—it’s a feature. Between Doximity marketing a “free 99” scribe, the OpenAI x Penda Health study showing that workflow integration mattered more than model complexity, and the recent discourse suggesting prompting > fine-tuning—I can’t help but think we will inevitably enter a contraction phase. I’m especially watching the SMB and middle market space, where dozens of ambient scribe offerings have popped up. My hunch is that many won’t survive the next few years—not because scribing isn’t valuable, but because ambient documentation alone may not be the moat we once thought. Still, I’m open to being wrong. Maybe scribe is less Pets.com during the dotcom bust and more like the rails that the ecosystem of AI care tools will build upon. 📉 Or 🛤️? I shared some brief reflections in a short piece here on The Techy Surgeon Substack. Would love your thoughts! https://lnkd.in/eFwgKATY

179

45 Comments

Great article describing the differences between API vs MCP. There are several tools out there (including open source) to help with the conversion - but in a large company with 1,000s of APIs, how much will it cost to convert given the dependencies? Is your API catalog ready (please say you have one...)? https://lnkd.in/e597cTq3

26

1 Comment

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