关于
Results oriented leader with 12+ years of experience in Healthcare, Data Science, AI…
服务
Sravan K.的文章
动态
2816 位关注者
-
Sravan K. Elineni 发布了此动态I've architected AI systems across healthcare, telecom, pharma, and robotics. One lesson keeps repeating. Don't start with the model. Start with your Gold Semantic Layer. In traditional ML we followed CRISP-DM. Understand the business, understand the data, prepare, model, evaluate, deploy. That still works for classical ML. For AI it's different. The foundation isn't a clean dataset. It's a Gold Semantic Layer - a curated, validated, business-defined layer where every field has a clear meaning, every relationship is documented, and every metric is agreed upon before any model touches it. Without GSL, your RAG system retrieves the wrong context. Your agent makes decisions on ambiguous data. Your multi-modal model fuses signals that mean different things in different departments. Garbage in, confident garbage out. Here's the stack I prioritize now. In order. Gold Semantic Layer first. Define what every data element means across the business. Revenue, patient encounter, dispatch, churn - these words mean different things to different teams. Align them once. This becomes the single source of truth that every AI system downstream inherits. Abstraction layer second. This is where business knowledge lives. Semantic memory that defines how your organization thinks - your entities, relationships, context. And business process memory that captures how work actually gets done - the workflows, decision rules, edge cases your teams have learned over years. When these are abstracted properly, any model can plug in and operate with your business context without being hardcoded to one provider. Evaluation layer third. This is where AI evaluation is fundamentally different from ML. Traditional metrics still apply at the model level. But for AI systems in production, you need judge models evaluating outputs against your GSL definitions. Eval agents that test reasoning chains, catch hallucinations, and validate that the system is actually following business process memory. Pair automated judges with human review and you get an evaluation layer that scales without losing accountability. Governance layer fourth. Who has access. What data flows where. How decisions get explained. How to roll back etc. Model selection last. Gemini 3.5 dropped last week. GPT-5.5 and Claude Opus 4.7 dropped the month before. Three frontier models in six weeks. If the GSL and architecture are right, evaluate the new model on Tuesday, swap it in on Wednesday, and nobody downstream notices. we should always have model zoo and let agents pick the best and available models. one for planning, 1-2 as work horse models and few as eval models. The model is the least important decision in your AI stack. The Gold Semantic Layer is the most important. Does your organization have a GSL or similar semantic or context layer? Or is every team defining their own version of the truth? #AI #EnterpriseAI #MachineLearning #AIArchitecture #DataScience #DataGovernance
-
Sravan K. Elineni 发布了此动态I didn't know Soma Somasegar personally. I wish I had. He passed away on Monday at 59. And my entire LinkedIn feed became something I've never seen before. Hundreds of people. Same message. Different words. "He believed in me before I believed in myself." Not one person talked about his titles. Not the 27 years at Microsoft. Not leading the Developer Division that built Visual Studio and .NET. Not the Madrona investments. Not the SEED 100 list he was named to hours before he died. Every single post was about how he made people feel. He remembered names. He made introductions without being asked. He championed people when they weren't in the room. He gave his time to founders who had nothing to offer him in return. Satya Nadella, Sriram Krishnan, UiPath, Madrona, hundreds of engineers and founders all saying the same thing. He made everyone around him grow. That hit me hard. I've been building my career in AI/ML for over a decade. I've focused on my technical depth, complex architectures, shipping production systems. But reading these tributes made me ask myself a question I don't ask enough. When people talk about me one day, what will they say? Did i uplift others around me? I wish to aspire to be a beacon like Soma one day. I want them to say what they're saying about Soma. At least i will aspire to be.. Rest in peace Soma. What's the best career advice someone gave you that had nothing to do with technology? #RIPSoma #Leadership #Mentorship #TechCommunity
-
Sravan K. Elineni 发布了此动态Google I/O just wrapped. Here's what nobody is talking about yet. Everyone is covering Gemini 3.5. New model. Better benchmarks. Expected. The real announcement was Antigravity. Google's agent-first development platform. Think about what just happened in the last 30 days. Anthropic launched 10 financial agents. Negotiating a $900B valuation. 80x enterprise growth. IBM shipped watsonx Orchestrate for multi-agent systems at Think 2026. OpenAI partnered with PwC on finance agents. And now Google drops an entire platform purpose-built for orchestrating agents. These companies aren't competing on models anymore. They're competing on who owns the agent layer. A migration agent in Android Studio that converts React Native or iOS apps to Kotlin in hours instead of weeks. That's not a chatbot. That's an agent that reads a foreign codebase, understands the architecture, and rewrites it. Last week Salesforce said context engineering is replacing prompt engineering. This week Google proved it. Antigravity isn't about better prompts. It's about giving agents the right tools, the right data access, and the right orchestration to actually execute. The model race is becoming table stakes. The orchestration race is where the value is moving. Interesting shift. What's catching your attention from I/O? #GoogleIO #AI #Gemini #AgenticAI #EnterpriseAI #MachineLearning
-
Sravan K. Elineni 分享了此动态Most AI projects don't fail because the model was wrong. They fail because the model was unnecessary. I've seen teams spend months building an LLM pipeline for something a regex could handle in an afternoon. Watched an agent framework get deployed for a task that needed a SQL query and an if-statement. Before reaching for an LLM or an agent, there's a decision ladder worth walking through first. Can a rule or a regex solve this? Can classical ML solve this? Can RAG solve this? Can a fixed 2-step workflow solve this? Does this actually need a custom multi-modal model? Only then - does this need an agent? Each rung has a cost, a complexity ceiling, and a failure mode. Skipping rungs doesn't make you innovative. It makes your project fragile. I put together a framework that walks through this ladder, when a custom multi-modal model is genuinely justified, and what it takes to run agents safely in production. I'm calling that last part harness engineering - because agents without guardrails aren't autonomous, they're unaccounted for. 15 slides. No fluff. PDF attached. Where on the ladder does your current project actually sit? #AI #MachineLearning #EnterpriseAI #AgenticAI #DataScience
-
Sravan K. Elineni 发布了此动态AI Agent Conference just wrapped in Manhattan. The stat that matters. 79% of organizations report AI agent adoption. Only 11% are running agents in production. Meanwhile this week alone IBM shipped watsonx Orchestrate, Anthropic dropped 10 financial agents, Coupa launched agent orchestration for supply chain. Everyone is shipping frameworks. Almost nobody is shipping agents that actually work. The single-agent problem is solved. One agent, One task, One Skill.md file. Done. The multi-agent problem is different. Agents talking to agents. MCP for tool connections. A2A for agent-to-agent communication. Deciding which agent handles what. Tracing a decision back through four agents to explain why it happened. Salesforce said it at the conference and I agree. Context engineering is replacing prompt engineering. Agent performance depends less on how you ask the question and more on the information architecture surrounding the agent. Which data it sees. How context is structured. What gets retrieved and when. We're entering the era of the agent architect. Not the person who builds agents. The person who makes agents work together. Is your org in the 79% or the 11%? #AI #AgenticAI #EnterpriseAI #MCP #MachineLearning #AIAgents
-
Sravan K. Elineni 发布了此动态Everyone is excited about OpenAI's Privacy Filter for scrubbing PII. And they should be. 96% F1 score, runs locally, Apache 2.0 license. But nobody is asking the harder question. How do you know your dataset is actually clean after scrubbing? How do you validate that you didn't destroy the signal while removing the sensitivity? Most teams scrub the data, hand it off, and hope for the best. There's a technique that solves this and almost nobody in the AI community talks about it. Generalized Low Rank Models. GLRM decomposes your dataset into low-rank representations that work across mixed data types simultaneously. Numerical vitals, categorical codes, ordinal severity scores, boolean flags. All in one framework. Not just PCA for numbers. GLRM handles the heterogeneous mess that real clinical and enterprise data actually looks like. Here's how I use it for validating privacy-scrubbed datasets: 1) Run GLRM on your original dataset. 2) Learn the archetypes and fix them. 3) Now project your scrubbed dataset onto those same archetypes and measure the reconstruction error and distance. Small distance means your scrubbing preserved the analytical value. Reconstruction error spikes means you over-redacted and your downstream models will suffer. Not "we scrubbed PII and it looks fine." Instead "we projected the scrubbed data onto the original archetypes and reconstruction error increased by only 2.3%, confirming analytical integrity." That's the difference between compliance theater and actual privacy engineering. Five questions to ask before trusting any scrubbed dataset: Did you measure reconstruction error before and after? Did you validate across all data types or just numeric columns? Can your model still detect the rare conditions after redaction? Did scrubbing create new correlations that weren't in the original? Would you trust a clinical decision made on this scrubbed data? If you can't answer these, your privacy pipeline has a measurement gap. What's your approach for validating data quality after PII and PHI removal? #AI #Privacy #MachineLearning #HealthcareAI #DataScience #GLRM
-
Sravan K. Elineni 发布了此动态Privacy and Security in healthcare: OpenAI just open-sourced Privacy Filter. A 1.5B parameter model that runs locally on your laptop and scrubs PII before it ever hits the cloud. Names, addresses, emails, phone numbers, account numbers, API keys. 96% F1 score out of the box. This is a big deal. And it's free. Apache 2.0 license. But here's what I immediately noticed. It covers PII. It does not cover PHI. In healthcare, PII is table stakes. The real challenge is Protected Health Information. Diagnosis codes. Medication lists. Lab results tied to a patient. Radiology reports. Clinical notes describing a specific condition for a specific person. HIPAA doesn't care that you scrubbed someone's name if their clinical note still says "45-year-old male presenting with HbA1c of 11.2 and bilateral retinopathy at Memorial Hospital on March 3rd." That's identifiable without the name. I've spent a lot of time on this problem. Built privacy-preserving multi-modal models using split learning specifically so patient data never leaves the hospital environment. The whole architecture was designed around the idea that you can train on clinical data without centralizing it but when you use clinical notes, i am always paranoid if we have done enough to scrub PHI eventhough other privacy guaratee exist. some transformer based models exists now but no where near 96% f1score. OpenAI's Privacy Filter is an excellent foundation. But for healthcare AI to move forward safely, we need to build on top of it. PHI-aware scrubbing that understands clinical context. That knows "Memorial Hospital on March 3rd" combined with a rare diagnosis is just as identifying as a name. That can redact at the semantic level, not just the pattern level. The model architecture is there. The 128K context window handles full clinical documents in one pass. The MoE design keeps it fast enough for real-time pipelines. Someone just needs to fine-tune it for healthcare. Who's working on this? or should i work on it? #HealthcareAI #Privacy #HIPAA #OpenAI #MachineLearning #ClinicalAI
-
Sravan K. Elineni 发布了此动态Anthropic just launched 10 AI agents for financial services. Today. At an event with Jamie Dimon sitting in the front row. These agents build pitchbooks. Review earnings transcripts. Run KYC checks. Reconcile general ledgers. Close month-end books. Audit financial statements. FactSet dropped 8% within hours of the announcement. Morningstar erased its gains. S&P Global and Moody's both sold off. Wall Street is watching its workflow get automated in real time and pricing it in the same afternoon. But here's what caught my attention. Dario Amodei said Anthropic projected 10x enterprise growth. They got 80x instead. That's not a typo. Eighty times what they planned for. Yesterday it was Anthropic with Goldman Sachs and Blackstone forming a $1.5B joint venture. Today it's 10 purpose-built agents. OpenAI countered with a PwC partnership for finance agents on the same day. Two years ago these companies were competing on benchmarks. Now they're competing for CFO budgets. For anyone in financial services, this is the moment to ask. If AI agents can close your books and audit your statements, what does your team's role become? or is this fluff? comment below what you think What part of your workflow would you hand to an AI agent tomorrow if you could? #AI #FinancialServices #Anthropic #EnterpriseAI #MachineLearning #WallStreet
-
Sravan K. Elineni 发布了此动态IBM just surveyed 2,000 CEOs globally. The results dropped today. By 2030, CEOs expect 48% of operational decisions to be made by AI without human intervention up from 25% today. But here's the number that stopped me. Only 25% of the workforce is using AI regularly right now. Despite 86% of CEOs believing their employees already have the skills. Read that again. The CEOs think the people are ready. The people aren't using it. That's not a skills problem. That's a trust and access problem. 83% of the same CEOs said AI success depends more on people's adoption than on the technology itself. They know this. And yet the investment still goes to models and infrastructure, not to helping the people actually use them. Last week I posted about the support gap. How IT colleagues are being told to "adopt AI" without anyone explaining what that means. This data confirms it at scale. The technology is moving. GPT-5.5 dropped last week. Claude Opus 4.7 the week before. Four Chinese labs released competing models in a 12-day window. The models are not the bottleneck. The bottleneck is the 75% of your workforce watching from the sidelines because nobody built the bridge for them. 29% of employees will need reskilling for a different role by 2028. 53% will need upskilling in their current role. That's not a future problem. That starts now. Who's building the bridge at your organization? Or is everyone still waiting for someone else to go first? #AI #EnterpriseAI #Leadership #FutureOfWork #MachineLearning
-
Sravan K. Elineni 赞了此动态Sravan K. Elineni 赞了此动态New free agentic AI courses coming soon. Follow for updates! USF Bellini College of Artificial Intelligence, Cybersecurity and Computing USF Innovative Education
-
Sravan K. Elineni 赞了此动态Sravan K. Elineni 赞了此动态It’s hard to find the words to write this, but my father, BRUCE, passed away on Friday. My dad’s “Big 3” in life were family, financial planning, and faith. He dedicated his life to all three, and I wouldn’t be where I am today without him. I wouldn’t be in financial planning if it wasn’t for my father. I wouldn’t have become a CFP® professional, and I certainly wouldn’t be in the position I am today without his sacrifices, guidance, and the foundation he built for me. There would be no Wise Capital Partners without Bruce Wise. I had the incredible privilege of following in my father’s footsteps, eventually taking over the business he started and building upon his legacy. That is something I will never take for granted. My dad went through so much over the last several years, but he never stopped being Bruce. On Friday, he was at home, eating a cheeseburger and fries, his favorite, and watching sports. Honestly, I can’t think of a more Bruce way to spend his final moments. As heartbroken as I am, I’m also happy for him. He is finally at peace and pain-free. I know he was excited to see his mother again, who he hadn’t seen in 50 years, and most importantly, I know he is with Jesus. Dad, thank you for every sacrifice you made for our family and for me. Thank you for believing in me, teaching me, challenging me, and giving me the opportunity to carry forward something you spent your life building. I promise we will continue your legacy and take Wise Capital Partners to the next level. I’m going to use your life, your strength, and your legacy as fuel and inspiration to keep moving forward. I know you wouldn’t want it any other way. Thank you to everyone who has shown our family so much love, support, and prayer during this time. Please continue to keep us in your prayers. I love you, Dad. I’ll miss you every single day, and I’ll spend the rest of my life making you proud. As Bruce would always say: “Have a wonderful life.”
-
Sravan K. Elineni 赞了此动态Sravan K. Elineni 赞了此动态ACL day #3: Attending poster sessions is strictly superior to oral presentations, and here's why: Watching someone talk on stage for twenty minutes is passive knowledge consumption, but standing by a poster gives you the opportunity to ask questions, actively interact with the knowledge, and fill out holes in your understanding. Actively consuming knowledge better internalizes it than passively consuming it! You get to stress-test your ideas, uncover blind spots, and most importantly - build real human connections. It is simply more effective, and honestly, it is a lot more fun. To those of you who present on stage, just remember that the burden is entirely on you to make it land. Since you do not have the safety net of back-and-forth interaction, you have to be obsessive about how you package the narrative. It is not about making it complicated to show you are smart. You have already been accepted! Now it is time to make it clear and simple, so everybody can actually enjoy the fruits of your hard work 🌟
-
Sravan K. Elineni 赞了此动态Sravan K. Elineni 赞了此动态It was an honor serving this great nation in uniform for 25 years, and I continue to try to give back. Happy 250th Independence Day! Let's make the future better.
-
Sravan K. Elineni 赞了此动态Sravan K. Elineni 赞了此动态What if reimagining work starts with stopping? Not adding another meeting. Not creating another process. Not making teams move faster without helping them focus better. During Novartis Culture Leadership Advisory Board, I stopped Rita McGrath, Lisa Bodell, and Muneto Mogi to ask them what they have stopped doing to create more clarity, energy and impact in an AI enabled world. Their perspectives are a reminder that culture is shaped through everyday choices: what we simplify, what we prioritize and what we make space for. Curious what they said? Watch the video and think about one small thing your team could stop, simplify or do differently this week. Small shifts. Big impact. #CLAB #SmallShifts
-
Sravan K. Elineni 赞了此动态Sravan K. Elineni 赞了此动态I’m thrilled to be starting a new role as a Senior Recruiter for Optum Tech! After spending nearly 11 years with UnitedHealth Group, returning to the Optum family feels both familiar and incredibly energizing. I’m also grateful for the opportunity to get back to my Talent Acquisition roots connecting with talented individuals and helping build teams that make a real impact. I’m looking forward to collaborating with amazing people, learning, growing, and contributing to the innovative work happening at Optum Tech. Thank you to SO many that helped get me to this point and to Zach Tufte for the opportunity. Here’s to new beginnings (that feels a lot like coming home)!
工作经历和教育背景
-
Brightspeed
查看Sravan K.的完整工作经历
查看他们的职位头衔、任职时间等。
或
资格认证
志愿者经历
出版作品
-
Optimizing care gaps and outreach programs in Healthcare
查看作品- defines some basic examples where data science can be established to reduce expenditure and improve efficiency in healthcare
专利发明
语言能力
-
English
母语或精通双语
-
Telugu
母语或精通双语
参与组织
-
CARD - center for autism related disabilities
mentor
收到的推荐信
14 位会员推荐了Sravan K.
加入领英，即可查看