Abhishek Srivastava

Stanford and Harvard released their State of Clinical AI report for 2026. One finding stood out to me. When AI systems were tested in conditions resembling real clinical work — incomplete data, ambiguous results, decisions requiring revision — performance dropped sharply. On reasoning under uncertainty, models performed closer to medical students than experienced physicians. Most people read this as an AI problem. I read it as a data infrastructure problem. Incomplete information doesn't appear from nowhere. It comes from pipelines that drop records silently. From lab results that arrive without context. From data that was modified, corrected, or overwritten somewhere between collection and model input — without a trace. In regulated clinical environments, we have a principle for this: you must be able to answer who created this data, who touched it, and whether it's still what it was originally. Most systems feeding clinical AI today cannot answer those questions. The model isn't failing because it's weak. Sometimes it's failing because what it received was already broken before it arrived. Clinical AI sits on top of data infrastructure that, in most hospitals and research environments, was never designed to support it. That gap is what I work on. Full report via Stanford Department of Medicine — link in first comment. #ClinicalAI #HealthcareData

18

1 Comment

🚀 Introducing Dicomaster (Available in GitHub & PyPI) Secure, Fast, Built for clinicians, researchers, and AI developers. “Every byte of data represents a life worth protecting.” — Santo Paul When I worked on medical imaging challenges like Radiological Society of North America (RSNA)’s Intracranial Aneurysm Detection, I realized one big problem, that is processing DICOM files safely, quickly, and without risking patient data was a nightmare. So I built Dicomaster: a high-performance, privacy-first DICOM toolkit that extracts metadata, anonymizes PHI, and preps medical imaging datasets for AI research, Kaggle challenges, and hospital systems. ⚙️ What is DICOM and what does Dicomaster do? -Have you ever wondered how medical images like MRI, CT, or X-ray scans are stored and shared? Modern imaging machines don’t just capture pictures, they also store patient details, such as name, age, scan descriptions, and physician notes. All this information is saved as metadata inside specialized medical file formats like .dcm or .ima, collectively known as DICOM (Digital Imaging and Communications in Medicine). -The problem? These files are complex and not easy for regular users or even many physicians to interact with. And when developing modern AI models to detect diseases (like brain aneurysms, breast cancer, or lung conditions) researchers need access to large datasets of medical images. However, there’s a huge challenge: Patient data must remain confidential. If we don’t anonymize or properly secure this information, we risk violating regulations like HIPAA and compromising patient privacy. That’s where Dicomaster comes in. It’s a secure, high-performance tool that helps you: - Anonymize DICOM files (remove or pseudonymize personal data) - Organize and classify massive datasets for AI or research - Process thousands of images in seconds - Detect urgency (for example, flag scans that show possible aneurysms or internal bleeding) - Convert .dcm/.ima files to another image file formats Read about this on: Medium: https://lnkd.in/eaHn5YHd Github: https://lnkd.in/e-NU8uRg #Dicomaster #MedicalAI #HealthcareInnovation #RadiologyAI #DICOM #MedicalImaging #DataPrivacy #AIForGood #MachineLearning #DeepLearning #FastAI #PythonDevelopers #OpenSourceProject #HealthTech #AIHealthcare #PatientSafety #DigitalHealth #DataScience #AICommunity #Innovation #Kaggle #AIMed #ClinicalAI #ResearchTools #Neuroscience #MedicalSoftware #Anonymization #HIPAA #GDPR #TechForGood

15

1 Comment

I was speaking with Chirag Patel, who leads labeling and inference at Intuit, about how to responsibly integrate AI coding assistants like Claude, Copilot, and Cline into development lifecycles in regulated industries such as healthcare and financial services. AI coding assistants are improving rapidly and can meaningfully accelerate development. But in regulated environments, acceleration without governance introduces compliance, security, and audit risks that organizations cannot afford. Three risks stand out. First, security and code quality. Research from Stanford and NYU found that developers using AI coding assistants were more likely to introduce security vulnerabilities and were often more confident their insecure code was safe. Studies on Copilot have also shown insecure pattern reproduction and vulnerable dependency inclusion when outputs are not carefully reviewed. The danger is not just flawed code. It is misplaced confidence. Second, compliance and documentation. Under HIPAA, SOC 2, PCI-DSS, or SOX, teams must demonstrate why code was introduced and how it satisfies governance requirements. AI assistants do not generate defensible audit trails or justify decisions against regulatory controls. That gap must be closed through engineered process. Third, sensitive data exposure. Sending code that touches PHI, PII, or financial logic to external models introduces data residency and contractual risk. Many enterprise agreements require explicit data processing terms. Without them, routine usage can create compliance exposure. The response should be governance by design. AI-generated code should be validated independently of the developer workspace before it reaches the main branch. Deterministic compliance gates in CI/CD must enforce security scanning, license validation, and domain-specific checks. If a change violates policy, it should not merge. Organizational policy should be expressed as code. Tools such as Open Policy Agent and Checkov allow regulatory controls to be machine-enforced and auditable. Compliance moves from documentation after the fact to enforcement at the point of change. Human oversight remains mandatory, especially for changes affecting data access layers or infrastructure. Given the tendency to over-trust AI output, review discipline must increase, not relax. Every AI-generated line should be treated as untrusted input. That framing strengthens review rigor. Static checks at merge are not enough. Runtime monitoring and anomaly detection are necessary to detect compliance drift in production, particularly when systems touch real patient or financial data. In regulated industries, teams must show reasoning behind changes, prove alignment with security standards, and maintain a traceable audit trail. AI coding assistants can improve productivity. But in healthcare and finance, that productivity is sustainable only when it is wrapped in governance that makes AI contributions safe and audit-ready.

13

Excited to see our effort being accepted at an #ICLR 2026 workshop. I had opportunity to mentor Yahia Alqurnawi , Preetom Kumar Biswas , and Anmol Rao on this work through the #CSE 576 #NLP course project at Arizona State University, and this being their #first research project makes the milestone even more special. Our paper, ViTaB-A, evaluates whether #multimodal #LLMs can not only answer questions over #tables, but also correctly localize the supporting #evidence at the row, column, and cell level across #Markdown, #JSON, and table #images. The results highlight a major gap between answer accuracy and #trustworthy #attribution, with attribution especially weak in JSON, better in images, and generally easier for rows than columns. The paper also examines whether model confidence aligns with attribution quality, and shows that this #alignment remains #unreliable. It was fun working with them, and I’m sure this is only the beginning. Many more to come. Finally, I would thank Prof. Chitta Baral and Dr. Vivek Gupta for there support. School of Computing and Augmented Intelligence #NLP, #TrustworthyAI, #ComputerVision

22

4 Comments

Fine-Tuning LLMs? Here’s What No One Tells You 🚨 Fine-tuning a Large Language Model isn’t just about feeding it more data. It’s like coaching a gifted student — each mistake needs its own kind of correction. In his latest breakdown, Bhavishya Pandit shares one of the most practical and clear-headed guides on LLM fine-tuning I’ve come across. He tackles the real production issues that derail models: 🧠 Overfitting – when your model memorizes instead of learning 💥 Catastrophic Forgetting – when fine-tuning wipes out general intelligence 📉 Poor Data Quality – the silent killer of performance 💸 High Compute Costs – scaling that breaks the budget 📏 Evaluation Complexity – when metrics mislead 🔁 Deployment Drift – the decay that no one plans for Each issue comes with clear symptoms, when to solve, and actionable remedies — from LoRA to rehearsal strategies to validation best practices. #LLM #MachineLearning #GenAI #AIEngineering #AI #DeepLearning #MLOps #FineTuning #AIEthics #NLP #ArtificialIntelligence #BhavishyaPandit #LinkedInLearning

81

1 Comment