Luca Martial

Love this important new piece by OpenAI’s incoming CEO of Applications (and former Instacart CEO), Fidji Simo, about the ways in which AI will empower people. And I couldn’t agree more with her about the exciting breakthroughs for AI in healthcare. As Fidji writes, healthcare is also about what happens between doctor visits, and AI can empower people to improve their health by improving their key daily habits: “AI can also make sure that health decisions don’t just happen in the doctor’s office. The biggest levers for preventing disease and optimizing health outcomes—sleep, food, movement, stress management, connection—all depend on everyday habits. AI can help us build those habits through small, achievable, daily steps with personalized, real-time nudges. Beyond diagnostics and accelerated drug development, this is one of the most profound uses of AI in healthcare: an always-on companion that bridges knowledge gaps, reinforces good habits, and empowers people to take charge of their own health with confidence and clarity.” But as Fidji notes, realizing the ultimate possibilities of AI won’t happen on its own: “That’s why we have to be intentional about how we build and share these technologies so they lead to greater opportunity and prosperity for more people.” It’s great to have someone as visionary as Fidji helping to build the AI that can truly empower people to take charge of their health.  You can read the piece here: https://lnkd.in/dCNmw8Ex #Health #AI #Technology #ArtificialIntelligence #Healthcare

379

65 Comments

Boehringer Ingelheim partners with Tempus AI in Oncology Real-World Data Pipeline Push: 💊Boehringer Ingelheim has signed a multi-year deal with Tempus AI to enhance its oncology pipeline using real-world patient data and AI-driven analytics 💊 The deal gives Boehringer access to over 8 million de-identified patient records, covering clinical, molecular, and imaging data 💊 Tempus’ Lens platform lets researchers build patient cohorts, explore biomarkers, and generate hypotheses for novel targets and combination therapies 💊 An AI assistant, Tempus One, can automate cohort creation using inclusion/exclusion criteria, saving time and sharpening precision 💊 This deal builds on earlier work and reflects a broader shift toward AI-powered, data-rich cancer R&D, aiming to make treatment more personal, predictive and faster 💊 Boehringer aims to accelerate progress in hard-to-treat cancers like small-cell lung cancer, NSCLC, liposarcoma, and gastrointestinal cancers 💊 The company’s pipeline includes experimental agents like a DLL3 T-cell engager, a HER2 inhibitor, and an MDM2-p53 antagonist now in Phase II or III. 💊 Boehringer says layering real-world data onto preclinical datasets helps deepen understanding of cancer biology and boost research efficiency 💬Last month I reported on AstraZeneca partnering with Tempus AI in a Gen AI drug discovery project 👇Sources in comments #digitalhealth #pharma #ai

724

8 Comments

Proud to share our first research paper: we've completed the first large-scale validation of an autonomous AI doctor in real clinical practice. Our study compared Doctronic's AI with board-certified clinicians across 500 urgent care encounters. The results? 81% diagnostic concordance and 99.2% treatment plan alignment, with zero clinical hallucinations. What struck me most: when experts reviewed discordant cases, AI outperformed humans ≈36% of the time versus ≈9% the other way. Not because AI replaces clinical judgment, but because it synthesizes vast medical knowledge consistently. This is a preprint but it seems that at Doctronic we're on the right path to address the projected shortage of 11 million healthcare practitioners by 2030. Technology that raises the floor of healthcare isn't just possible - it's here. Read the full study: https://lnkd.in/e3vUhsNs #HealthTech #DigitalHealth

108

11 Comments

We’re excited to collaborate with MongoDB and CentralReach on one of the most concrete examples of AI agents in healthcare: Transforming vast amounts of unstructured client and treatment data, extracting out both semantic and tabular insights, and using it to fuel a multitude of “agentic” downstream tasks: * A comprehensive overview of a patient (Care360)  * Decision support and recommendations * Longer-term care planning All of these are agentic workflows built upon a central knowledge base with LlamaCloud + MongoDB. Check it out: https://lnkd.in/gkpe82zw If you’re interested in signing up for LlamaCloud with our MongoDB integration, signup to LlamaCloud: https://lnkd.in/g9Wpqn7w Contact us: https://lnkd.in/gE9aSVV7

143

6 Comments

Excited to announce JustPaid's partnership with Scritch (YC W24). Claire Lee and Rachel Lee are building an AI receptionist for veterinary practices – handling calls 24/7, managing appointments, and answering questions with real veterinary expertise. Vet professionals are burning out while pet ownership grows. Scritch is building AI coworkers to lighten their load and help more pets get care. JustPaid will handle all their billing so they can focus on what matters: revolutionizing veterinary care. Two Y Combinator AI-first companies working together to remove operational friction and enable real innovation.

36

15 Comments

Just-in-Time UI: At Aligned Labs we frequently share reports about the data collection, fine-tuning and eval work that we do with clients.�� They often need different views or refined analysis, which is why we built Whiz-bang Boom! Our new tool takes a CSV or JSON file and generates a shareable, interactive report. You can easily modify the report or focus the analysis with a simple prompt, and it creates a new revision. You can organize reports into spaces and customize those spaces with a simple prompt. Try it out. https://wizbangboom.com #businessIntelligence #AI

25

2 Comments

Most labs are sitting on a goldmine of underused reaction data. Few are making the most of it. High-throughput experimentation (HTE) can screen thousands of reaction conditions every month - but in most organizations, that data isn’t feeding back into future optimization efforts. Multitask Bayesian optimization (MTBO) is one way to change that - by learning from past experiments to make future ones smarter. Here’s the idea: Traditional Bayesian optimization treats every new reaction like it’s starting from scratch. But what if we could learn from earlier experiments - even if they involved different substrates or conditions? That’s where MTBO comes in. It draws on prior data to accelerate the optimization of new reactions with similar mechanisms, even when the match isn’t exact. This is particularly relevant when carrying out synthesis with commonly used named reactions, such as Amide couplings and Buchwald-Hartwig couplings. MTBO builds a prior from historical datasets, then learns how the old and new reactions are correlated. That relationship helps guide what to test next - getting to better results, faster. By building systems that learn within reaction classes, we can: → Optimize faster → Run fewer experiments → Get better outcomes with less data This is the future of reaction optimization: not just automation, and not just machine learning - but smart, adaptive platforms that improve over time. At ReactWise, we’re making that future real with multitask Bayesian optimization. If you’ve got years of reaction data sitting in ELNs or spreadsheets, let’s talk.

75

3 Comments

Today the Pulse team published a deep dive on why a single “accuracy” score doesn’t tell you if a document extraction system will survive in production. The goal here is to lay out an introductory but still rigorous evaluation methodology - we have an exciting open-source benchmark building on this research coming out very soon. Let’s do the math: take 1,000 pages, each with 200 data elements. A model that’s 98% “accurate” on paper still produces 4,000 incorrect values. Now make some of those: 1/ Broken reading order that scrambles multi-column layouts 2/ Tables with shifted columns or missing headers 3/ Cross-page context lost entirely That’s enough to silently corrupt an entire dataset without throwing a single error. We’ve processed hundreds of millions of pages and built a multi-axis evaluation framework to measure what actually matters: reading order validation, region-level ANLS, reading order accuracy, TEDS for table structure, and continuity checks across page boundaries. The result? Fewer silent data corruptions, more predictable performance, and pipelines that keep working on the next million documents you haven’t seen yet. Full technical write up in the comments!

36

2 Comments

Everyone is talking about the new DeepSeek-OCR model that dropped this week. Here's how it does on challenging, real-world documents: We're running multiple internal benchmarks, but here's an anecdotal example on a (publicly available) healthcare document. Unfortunately, the outputs leave a lot to be desired. - entire columns of data were dropped - signature blocks were often partially captured or dropped entirely - results were highly inconsistent (re-runs produce different outputs) It's why we built a hybrid OCR + VLM engine at Extend for parsing documents. Using OCR models for what they're really good at, and looping in VLMs for corrections, edits, and hydrating context as needed. That said, this release by DeepSeek is still a *huge* technical breakthrough. The point wasn't "we built a SOTA OCR model". Instead, the breakthrough is around around vision-tokenization / optical compression of document content. Traditional OCR models work by first transcribing the page into text and then converting that text into tokens, roughly one token per word, which are fed into a language model. DeepSeek’s approach flips this entirely. Instead of working in text space, it divides the original document image into visual patches and represents each patch as a vision token. By encoding visual information DeepSeek achieves ~10x compression while maintaining 97% accuracy of traditional text-based models. It’s an incredibly impressive piece of open-source work from the DeepSeek team. Excited to see where it goes from here.

66

3 Comments

We just open-sourced an eval harness for retrieval — built for the community. Accurate retrieval is hard to build. Not because embeddings are complicated — but because there's no easy way to test what's actually working. Run a sweep across chunking strategies, embedding models, and search parameters. Find what actually moves your recall. What's your current approach to evaluating retrieval? Vibes-based testing or something more rigorous? Link to the repo in the comments 👇 🎁 Day 4 of 12

59

1 Comment

Assort Health raises $76M Series B for their healthcare voice AI platform. Assort just raised their Series A a few months ago. The quick Series B is not to cover burn, but because they are succeeding so quickly.

44

1 Comment

The most spooky thing about AI is you can flip a coin for accuracy ☠️ The benchmark crisis is real with current metrics not telling you if a model is actually safe to deploy. MedPI is different. We built a comprehensive evaluation framework with 54-63 clinical competencies (varies by specialty) that measures: safety discipline, diagnostic reasoning, urgency recognition, interaction efficiency, empathy, adaptive dialogue. This means measuring a big umbrella metric like "safety" without breaking it down into specifics (e.g., medication errors) only tells half the story. To truly improve, you need to dig deeper and figure out what part of the clinical process went wrong: Was it the medication decision itself or was it the model’s failure to capture all clinical details in prior conversation? Understanding these nuances is key to developing safer, more effective medical AI. Here's how we validated it: We tested 9 major language models: GPT-5, Claude Opus 4, Claude Sonnet 4, Gemini 2.5 Pro, MedGemma, LLaMA 3.3, and others across 7,010 multi-turn patient conversations. Real clinical workflows such as general practice, dermatology, oncology, internal medicine, neurology, psychiatry, OB-GYN. And the results were surprising. GPT-5 topped the leaderboard at 53% accuracy - the best among them all and still being terribly inaccurate. Tf you mean, that’s crazy? Just flip a coin, and you’ll get the same level of accuracy. We're making this available because the industry needs a better standard, and MedPI Eval sets a new benchmark for understanding where medical AI stands today and how far it still has to go. Full whitepaper in the comments.

59

7 Comments