Felix Sanchez Garcia

The Dewey Data platform now has Attribute Mapping 🗺️ Academic researchers face a common challenge: working across datasets means navigating hundreds of different attributes, each with their own naming conventions and structures. Finding the common threads? That's been manual, time-consuming work. We've solved this for Dewey users. We identified the attributes that appear most frequently across our platform and mapped them to every relevant dataset. Now you can instantly see how datasets relate to each other based on shared attributes. No more guessing which datasets can be combined. No more spending hours cross-referencing field names. Check out the mapped attributes section in our docs to see it in action: https://lnkd.in/g2DUsX3h

18

1 Comment

A very brief, somewhat technical take on Large Language Models (LLMs). They are *effectively* analogs of Bloom filters for searching vector databases using natural language interfaces, and reporting back using natural language. This statement gives you all you need to know about potential and pitfalls, as long as you abstract the keywords to their essentials. Vector Database = encodes information using more or less artificial, possibly statistical descriptions of recorded facts about the objects in the database. Natural language = imprecise language ie cannot have formal proofs about internal consistency or ambiguity of the query or the query result. Bloom filter = sensitive but not specific way to search vector databases. Designed to not have false negatives (i.e. not miss), but will generate false hits (akin to hallucinations). There are also technical issues that make LLMs dissimilar to Bloom filters, but they win on using natural language as a query language; it frees people from formulating a very complex query in a formal language, reducing the barrier of asking questions for non-experts in a field. Furthermore, the answer is also not formulated in a formal language or served in a technical manner, even further reducing the barriers in interpreting answers. But here lies the problem: if anyone can ask and have their questions asked, how do you deal with the false hits, the outright lies, or more technically the false positives without the technical background to judge the answer, fact check it, or incorporate in a real world context? A society of engineers would acknowledge this limitation and use LLMs as accelerants the way we use high temperature settings in simulated annealing (SA) global optimization schemes: as a quick way to generate an approximate answer, and then painstakingly (the "cooling scheme" in SA) refine the answer by making it more precise. Unfortunately, we are not a society of engineers, but a culture of Dunning Kruger susceptibles. Enjoy your dopamine fix from the slop now. #AI #LLM

23

2 Comments

Applications for the AI Growth Zones are open (key criteria summarised on a page). Application deadline by the end of May 2025 to be chosen as an AI Growth Zone in the summer of 2025. This is part of the UK government's (Department for Science, Innovation and Technology) efforts to accelerate efforts to build the infrastructure needed to support AI growth. AI Growth Zones will unlock investment in AI-enabled data centres and support infrastructure by improving access to power and providing planning support. This will help drive innovation, create high-skilled jobs, and strengthen the UK’s position as a leader in AI. #AIActionPlan #AIGrowthZones #AI

11

At PyData Hull, we’re interested in more than just tools, we care about how data systems actually behave in practice. From data pipelines and reproducible workflows to model evaluation, interpretability, and deployment, strong data work comes from understanding trade-offs, assumptions, and real-world constraints, not just following the latest framework or trend. Python gives us an incredibly rich ecosystem to work with, but real progress happens when people share lessons learned, challenge ideas thoughtfully, and build reliable systems together. That’s the kind of community we’re building in Hull; a space for students, researchers, and practitioners to exchange practical knowledge, explore applied problems, and develop robust, responsible approaches to data science and machine learning. If you’re working with Python, analytics, ML, or applied AI, or curious about how these systems work under the hood, you’re very welcome to join as we’ll be announcing our first meetup shortly! #PyDataHull #Python #DataEngineering #MachineLearning #OpenSource #TechCommunity

8

In the latest AI Data Chats, Jennifer Ding from the Encord team sits down with the Ben Burtenshaw to talk all things vibe testing — from building domain-specific datasets to agentic workflows and verifiable tasks. If you’re serious about crafting high-performing ML models, this one’s a must-watch. Tune in to hear why deeply exploring your data isn't just helpful — it's essential.

5

everyone wants a faster database ... https://lnkd.in/guMDU4MV Here's a summary: DBPlanBench: A Test Harness for LLM-based Query Optimization Researchers introduced DBPlanBench, a test harness that enables LLMs to optimize database query plans. Key contributions: - DBPlanBench: A novel test harness with a token-efficient serialization format and patch-based editing interface for LLMs to refine execution plans. - LLM Optimization: LLMs outperform DataFusion optimizer in query plan optimization. - Semantic Cardinality Estimation: LLMs use domain knowledge to identify structural optimizations (e.g., join reordering). - Small-to-Large Workflow: Optimizations at small scales transfer to larger scales via rule-based procedures. Code available on GitHub 😊

4

1 Comment

At the end of last year, we started working with GenomeKey on two key hires - a Senior Data Engineer and an ML Engineer. Both roles came with some very specific technical asks. The ML search in particular went down an interesting route, with sourcing centred around mapping relevant publications and reaching out directly to researchers. The eventual hire was identified through her work published in Nature Portfolio. Both hires are now in place, and we've received some brilliant feedback from the founding team and the hires themselves. For similar requirements Joe Phillips is your go-to in BioAI. You can read more about his delivery to this search over on the website. 🔗 https://lnkd.in/ewNFkR25 #casestudy #dataengineering #mlengineering

8

2 Comments

An investigation of Meta’s AI as a therapist, the Prime Minister’s support of diagnostic AI, and parts of the UK’s new data law kicking in. Here is a news round up so you can focus on the real work on the ground. 𝐔𝐊 𝐇𝐞𝐚𝐥𝐭𝐡𝐜𝐚𝐫𝐞 𝐀𝐈 𝐍𝐞𝐰𝐬: 🇬🇧 This week the Prime Minister launched an AI Exemplar Programme which covers multiple sectors. In health there are two projects: AI diagnostics and AI discharge. Link: https://lnkd.in/eZRpQEH7 🇬🇧 The Guy's and St Thomas' NHS Foundation Trust Clinical Scientific Computing team have open-sourced their Quality Management System for the development of AI as a Medical Device. Well done to Dijana, Anil and the whole team. Link: https://lnkd.in/e3cscpFJ 𝐎𝐭𝐡𝐞𝐫 𝐫𝐞𝐥𝐞𝐯𝐚𝐧𝐭 𝐔𝐊 𝐀𝐈 𝐍𝐞𝐰𝐬: 🇬🇧 The first bits of the new UK data law, the Data Use and Access Act, came into effect on Wednesday. Here is a good summary by Robert Bateman (6 mins): https://lnkd.in/e8_7nyWX 𝐈𝐧𝐭𝐞𝐫𝐧𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐀𝐈 𝐧𝐞𝐰𝐬: 🇺🇸 Healthcare workers in Kaiser Permanente protested for a new contract that includes job protection against AI. Link: https://lnkd.in/eh_X3td7 🇺🇸 𝘛𝘳𝘪𝘨𝘨𝘦𝘳 𝘸𝘢𝘳𝘯𝘪𝘯𝘨: 𝘴𝘶𝘪𝘤𝘪𝘥𝘦. A parent who lost their daughter to suicide whilst they sought help from a chatbot, highlights the hidden space it creates can remove standard  required protections (like being involuntarily committed). A heartbreaking read. Link: https://lnkd.in/eJpGE2NU  🇺🇸 Meta and Character. ai are being investigated by the Texas attorney-general regarding if they misleadingly market their AI as therapists. Link: https://lnkd.in/eD3yzdNF

38

2025 was quite the year, but thanks to some down time, I’ve been able to catch up on recent work on AI and data ethics in the context of LMIC systems. I found the RAi UK white paper on Responsible AI in LMICs interesting for how it explores AI adoption in contexts where data systems are still immature and local priorities are overridden by global frameworks. It outlines the reality many of us see in health and social systems on the ground and raises questions about who innovation is really being designed for. UNESCO also updated the Playbook on tackling gender bias and harms in AI. I liked how practical this work is. It shows how the same inputs can still produce biased outputs. In settings where women’s data is poorly captured, these biases amplify inequalities and actively make them worse. Reading these back to back brought into focus the same clear message. Ethical AI is about about better models yes, but more so about data, the systems that produce it, and the decisions that shape its use. Adding AI does not improve decision making if data is incomplete and governance is patchy. It is just the same problems just dressed up as innovation. Curious about what 2026 will bring as this space keeps changing. 🔗 RAi UK White Paper https://lnkd.in/eiKNStaJ 🔗 UNESCO Red Teaming Playbook https://lnkd.in/ekW-M6u7

22

1 Comment

Excited to share our new preprint from the Advanced Forecasting team: “Forecast-Then-Optimize Deep Learning Methods” is now on arXiv: https://lnkd.in/gEzax9fM This work reflects our vision of building forecasting systems that not only predict—but also optimize real decisions. Key highlights: • A comprehensive review of deep learning–based FTO methods (2016–2025) • Practical guidelines, evaluation metrics, and taxonomy from real-world ops use cases • Roadmap for next-gen FTO systems, including the role of LLMs Proud to collaborate with Jinhang Jiang and Nan Wu. Thanks to Shivanku (Shiv) Misra for his vision in McKesson Enterprise Data Science Products.

15

The latest advancements in Retrieval Augmented Generation (RAG) are transforming how we interact with complex datasets, especially in scientific research. For R&D, this means more accurate insights, faster hypothesis generation, and a deeper understanding of interconnected knowledge. We're particularly excited about how optimized RAG pipelines can unlock value from vast academic literature for the next generation of our Research Compass tool, being tested now by our clients and expert network. What are your thoughts on RAG's potential in research? #AI #RAG #MachineLearning #DataScience #ResearchAndDevelopment #Innovation https://lnkd.in/eYNidzdp

7

How is AI changing our lives? 🎧 The Turing's Adrian Weller joined Brian Cox, Neil Lawrence, Steph Wright and Jeanette Winterson to explore this fundamental question. Listen to the latest episode of The Francis Crick Institute’s A Question of Science podcast: https://lnkd.in/eR3yedD3 #AI #AIEthics #AIforGood #AIGovernance

163

1 Comment

FunctionGemma, a new 270M parameter language model from Google, designed specifically for function calling and tool execution — not general chat.   Why is FunctionGemma special? 🔹 Extremely lightweight FunctionGemma can run in ~550MB RAM, making it possible to deploy on laptops, edge devices, and even phones — all without large GPUs. 🔹 Purpose-built for function calling Unlike general LLMs, this model is trained for text-only tool calling, making it ideal for automation, agents, and API-driven workflows. 🔹 Easy local deployment The model is released in GGUF format, so it works seamlessly with llama.cpp and Ollama using quantized versions (Q4/Q5/Q8). 🔹 Fine-tuning ready from day one Thanks to Unsloth’s day-zero support, you can: Run FunctionGemma on CPU or GPU Fine-tune it locally using Colab notebooks Choose full fine-tuning or LoRA   What can you fine-tune it for? FunctionGemma is meant to be customized for specific tasks, such as: Multi-turn function-calling agents Mobile actions (calendar events, reminders, flashlight, timers) Task-specific automation and workflows Lightweight, offline AI assistants This model clearly shows a shift toward small, focused, and efficient AI — where models are trained to do one thing well and can be adapted quickly to real-world use cases.   Model (GGUF): https://lnkd.in/gxVuBhZA Docs & fine-tuning guides: https://lnkd.in/gchpAjUF Run and Deploy LLMs on your iOS or AndroidPhone: https://lnkd.in/grk72Bey

5

I’m excited to share that our scoping review preprint has been published on medRxiv: A Scoping Review of Algorithmic Equity, Data Diversity, and Inclusive Design in the Transformer Era of Clinical NLP https://lnkd.in/eWvWM8h6 Many thanks to Abeer Badawi, Ph.D., Elham Dolatabadi, and Farah Ahmad, MBBS, MPH, PhD, for their collaboration, guidance, and support throughout this work. The rapid digitization of healthcare has positioned transformer-based NLP models as powerful tools for managing clinical text. However, their growing integration into practice raises important and unresolved questions about equity and inclusivity. This review synthesizes 56 studies (2017–2024) and examines how equity is addressed across three dimensions: ➡️ Algorithmic equity — fairness audits are largely post hoc and fragmented ➡️ Data diversity & representativeness — persistent underrepresentation creates what we define as Data Diversity Debt ➡️ Participatory design — observed in only 11% of studies, revealing a major gap beyond clinician involvement. To move beyond descriptive audits, we propose an equity-by-design roadmap to embed fairness, inclusivity, and accountability across the full lifecycle of clinical NLP systems—and to retire Data Diversity Debt. #HealthAI #ClinicalNLP #AlgorithmicEquity #DigitalHealthEquity #AIethics #ResponsibleAI #medRxiv

26

2 Comments

Rethinking tokenisation for genomic sequence data 🧬: our AI research Team at M42 Health has integrated biological context into the modelling process, significantly enhancing how we represent and analyse genomic information. Read some of the key highlights in the post below.

35

Super excited to be hosting our first BIO × AI Hackathon at Imperial as part of the UK AI Agents Hack Ep4! We’ve lined up two frontier tracks in Protein Engineering and Tissue System Analytics, designed for builders who want to work on real biological problems. If you're interested in AI applied to biology, this one’s for you. Hope to see you there!

20

4 Comments

⚡ What’s Actually Driving Innovation in RAG? It’s Not the Embeddings. In Episode 26 of LLMs on the Run, Professor Joseph Gonzalez explains why recent gains in retrieval aren’t coming from better embedding models—but from how we use them. Joey shares: ✅ Why embedding models are no longer the main bottleneck ✅ How chunking and query rewriting are driving real improvement ✅ What query expansion can do to improve context relevance RAG pipelines are evolving fast. Knowing what’s actually moving the needle can change how you build. Brought to you by 👟 LLM: https://lnkd.in/gna9guWX #LLMsOnTheRun #RAG #AIInfrastructure #RetrievalAugmentedGeneration #EmbeddingModels #AIResearch #RunLLM

6

In ADMET prediction, adding volume without complementarity can lead to stagnation or even degraded performance. In a new blog post, Lewis Mervin, Principal Machine Learning Engineer at Apheris, explores why data diversity and alignment matter more than volume when improving model accuracy, calibration, and applicability. 🧠 Why listen to Lewis? Lewis brings over a decade of experience in molecular property prediction, with a PhD in Chemistry from the University of Cambridge and six years at AstraZeneca, where he led efforts in ADMET modelling and evaluation. He played a central scientific role in MELLODDY, the largest cross-pharma federated ADMET initiative to date, working on its privacy-preserving modelling framework and collaborative design strategy. In this blog, he draws on: - Recent results from the ChemMedChem study by Schliephacke et al. - Learnings from our own Federated ADMET Network at Apheris 📌 The takeaway: complementarity, balance, and harmonization, not just scale, drive meaningful improvements in ADMET model performance. 👉 Read the full blog: https://lnkd.in/e_fZFtQA #ADMET #FederatedLearning #DrugDiscovery #MachineLearning

13