Vasilis Vryniotis

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. If you're an AI Founder in growth mode, reach out to my AI team at Cubiq Recruitment. Our specialist in BioAI is Joe Phillips. #Casestudy #AIscaleup #BioAI #DataEngineering

8

🚀 Benchmarking Local LLMs Like a Pro I just published a step-by-step guide on building a resumable evaluation pipeline for local Large Language Models (like Gemma 3 via Ollama). No more rerunning 10,000 examples if your script crashes—think save-game checkpoints for AI experiments. 🔗 Read the full blog here: If you’re working on LLM benchmarking, retrieval-augmented generation, or local inference workflows, I’d love to hear your thoughts! #LLM #MachineLearning #Python #Ollama #MLOps

8

2 Comments

I'm proud to announce that at PyMC Labs, we've published a quantitative, reproducible comparison between PyMC-Marketing and Google Meridian, two of the leading tools for Bayesian Marketing Mix Modelling (MMM). Most existing comparisons are feature-level and qualitative, leaving teams to choose based on brand name or comfort with a particular codebase. We wanted to go one step further. Together with my colleagues Teemu Säilynoja and Luca Fiaschi, PhD, we built a rigorous technical benchmark using realistic synthetic datasets that mimic everything from single-market startups to global enterprises. Check out the full blog post + code 👇

103

5 Comments

NotebookLM is an awesome tool for research and study. I think it is turning to be almost as a new interface for Search, we have been using Google as it is for years but sometimes our questions and what we are looking for are more specific or more complex than a websearch. Examples of usage: - Opening a coffee shop in Brazil -> It will give you many sources and information that you can filter or deep dive into. - Planning a trip - Learning a new skill This new way of addin sources expand our hability of using it as a "new Search" system. https://lnkd.in/g95dWSVe

15

1 Comment

Went through the Agentic AI MOOC from Berkeley RDI Something the course touched upon was Agent Evaluation. A lot of current evaluation being used for agents is essentially LLM evaluation; in this schema, the benchmarks only look at the final text output quality. This doesn’t fully capture the behavior of agents which have more complex behaviors such as tool usage, state and memory, and planning. Agents also live in a dynamic environment where the state may change, whereas LLM’s live in a static environment. To address these behaviors a shift from static benchmarks to dynamic benchmarks can better measure agent ability. Rather than having a dataset of inputs with deterministic, correct answers, dynamic benchmarks involve creating entire environments for agents to operate in. Within this environment, we can measure the decisions the agent makes at each step, tool usage, user interactions, and the resulting effects on the environment. We can then see how these decision led to the final result. The correctness of the agent’s decisions can be measured using human verifiers or “LLM-as-a-Judge” verifiers. With better evaluation benchmarks, researchers can make better-informed decisions on improving Agents. Link to MOOC: https://lnkd.in/eTfTnj6N

5

2 Comments

Ever noticed how RAG systems sound confident… and still get it wrong? That’s been my experience with vector-based RAG on long, professional documents. High semantic similarity, but low real relevance. Chunking breaks context, and retrieval feels hard to trust. Then I tried PageIndex — and it genuinely felt different. • PageIndex drops vectors entirely and retrieves information through structured reasoning, not similarity. • No vector DB • No chunking • LLMs reason over a document’s tree structure, like a human expert What stood out for me: • Clear section & page references • Much higher confidence in answers • Performance that matches real usage — 98.7% accuracy on FinanceBench Is it for every use case? No. Is it one of the best RAG systems today for serious, domain-heavy documents? Absolutely. Similarity ≠ relevance. Reasoning wins. #RAG #LLM #GenAI #AIEngineering #ExplainableAI #PageIndex

10

Great interview by David Hughes and the GraphGeeks crew! It's a strange coincidence that just a month before I joined LanceDB, a new open source project came about, lance-graph, that connects two worlds I've long been a part of. Interview: https://lnkd.in/gSK7zxsk lance-graph repo: https://lnkd.in/gBWmnj5a I've been closely following the lance-graph project over the last few months, and and the commits have been flowing, lately. Being built out in the open, and using the open source Lance format, it aims to bring graph querying capabilities to developers working with complex, multimodal datasets who may see the need for combining vector search with graph traversal. The query interface exposed is Cypher, which no doubt I and many other graph practitioners are very familiar with. The engineering of the query engine and index builder is also fascinating, and it's based on Apache DataFusion, a project I've long admired and always wanted to geek out on. Looking forward to exploring all of these concepts more and writing more about it soon. 🚀 In the meantime, check out the GitHub repo and give it a star! ⭐️ lance-graph repo: https://lnkd.in/gBWmnj5a

24

🧠 TOON: The Smarter, Lighter Alternative to JSON for LLMs Today I came across something that genuinely changes how we think about feeding structured data into Large Language Models: TOON (Token-Oriented Object Notation). If you’ve ever tried putting large JSON blocks into an LLM prompt, you already know the pain: 🔻 too many tokens 🔻 too much duplication 🔻 too many brackets 🔻 models getting confused by repeated keys JSON is great for machines, but it’s terrible for token efficiency. Every {, }, :, and repeated key costs extra tokens… and when you’re working with agents, multi-step reasoning, or large datasets inside a single prompt, that overhead matters. 🔹 What TOON Solves TOON restructures data in a compact, human-readable, LLM-friendly way that uses dramatically fewer tokens than JSON or YAML. Less noise leading to more reasoning space to get better model accuracy. It was introduced publicly on Nov 2, 2025, credited to Johann Schopplich and contributors. Built specifically for LLMs (not web APIs) and it optimizes structure so models interpret data more reliably. 🚀 Why This Matters In one of my own projects, the entire dataset fit into the context window, so we didn’t even need a RAG system. But the JSON made the prompt bloated and the model started: • misreading fields • losing relationships • confusing repeated keys The moment I saw TOON, it clicked that this might be the kind of efficiency we needed all along. 🔍 LLM Inputs = A Data Efficiency Problem The future of AI agents isn’t just about better models; it’s about how efficiently we feed them data. Formats like TOON open the door to lighter prompts, lower costs, and more reliable reasoning. Excited to experiment with this! If you’re working on LLMs, agents, or embedded AI systems, this is worth looking into. #AI #LLM #DataEfficiency #TOON #MachineLearning #ArtificialIntelligence #GenAI #PromptEngineering #AIInnovation #FutureOfAI #TechTrends #LearningEveryDay #ContinuousLearning #CuriosityDriven #GrowthMindset #AIDevelopment #LLMOptimization #TechCommunity #NewTools #ExploreAndLearn

11

I’ve been working on a new agent orchestrator called RobustNX, designed to solve the "black box" feeling of most AI tools. I wanted to build something that feels more like an IDE — transparent, stable, and collaborative. Here are three features I focused on to make it more usable: 1. User Memory Hub: A 5-Hub architecture that remembers context and preferences across sessions, rather than just basic chat history. 2. Live 'Thinking' Graphs: Instead of a loading spinner, it streams agent logic via SSE onto a dynamic graph, so you can watch the Planner and Coder work in real-time. 3. Built-in Circuit Breakers: Logic integrated into the core loop to catch infinite loops and prevent token-burn before they happen Check out the demo here: https://lnkd.in/guEKB9pK

23

I just watched a very interesting bauplan webinar on data quality with Claude Code https://lnkd.in/e6g9TM7U I highly recommend giving it a watch, a couple of things stood out to me: - the ability to use GenAI on actual data within a developer workflow, rather than in isolated chat tools - how the entire workflow stays within an IDE and terminal, including branch management

6

Data quality is everything. Without it you can forget AI, forget natural language assistants, forget attribution models, forget forecasting, forget machine learning, forget automated bidding. Without quality data you lose confidence in your insights and actions - and momentum from positive action and optimisation is impossible to build. Time is lost investigating data queries; time is wasted manually fixing data. What do I mean by quality data? 1. It’s accurate - the numbers you’re looking at reflect what really happened. 2. It’s meaningful - it describes what happened clearly and in a way you can act upon. How to achieve high-quality data? > Taxonomy is critical. A well-designed and properly maintained taxonomy transforms your data. Automated monitoring and exception flagging catch issues early, refine processes, and help all teams work in a consistent, reliable way. > Automated data pipelines refresh your data repeatedly, maintain the raw data that gives you accuracy, flexibility, and future-proofing. > A Semantic Layer that provides transparency and auditability, ensuring everyone knows what the data they’re looking at means - and trusts it. Flexibility to adapt and support new use cases is essential if users are to avoid reverting to the source data for answers. At Bright Analytics, we focus on delivering foundational, high-quality data that’s accessible in a transparent and flexible way - powering the use cases of all your teams. Our modular product and approach ensures you have high quality data by breaking up the task into manageable areas of focus: > have we got accurate raw data? (Automated Pipelines) > can we describe our data accurately? (Taxonomy) > do we understand our data models and its definitions? (Semantic Layer) My post from a while back about the poor, neglected data roadies rings true more than ever as the relentless messaging around AI and investment in the possibilities rolls on: the flashy stuff is only worth investing in if the fundamentals are in place. Invest in the boring fundamentals so you can truly leverage the exciting possibilities of emerging technology. #qa #foundations #automation #taxonomy #semanticlayer #dataroadies

12

1 Comment

"Without data, you’re just another person with an opinion." 🚗 With the rapid adoption of Electric Vehicles (EVs) in the UK, buyers face a big question: Which brand offers the best value? Which model balances price, efficiency, and performance? To answer this, I built a data-driven analysis project where I: 🔎 Scraped specifications of 700+ EVs from the UK market using Python & BeautifulSoup. 🧹 Cleaned the dataset (13 key features) by handling nulls, removing outliers, and standardizing values. 📊 Conducted Exploratory Data Analysis (EDA) to uncover trends in efficiency, range, weight, and pricing. 💡 Key Insights: ➡️ Top 10 Brands by number of models offered in the UK. ➡️Efficiency vs Weight analysis showing performance trade-offs. ➡️Price vs Range to highlight the best value models. ➡️Designed a new metric Price-per-Range (£/mile) to evaluate cost-effectiveness. ➡️Identified Top 5 brands/models that deliver maximum efficiency for money. ➡️Tesla, Audi, and BMW dominate the premium segment, while MG, Nissan, and Hyundai offer excellent value per mile. ➡️Vehicles with higher weight tend to be less energy efficient. ➡️Most EVs fall in the 200–300 mile range bracket. ➡️Acceleration (0–60 mph) strongly correlates with price. ➡️EVs with better efficiency are not always the most expensive. 📂 The project includes visualizations, statistical analysis (skewness, kurtosis, correlation, heatmaps) and is fully documented in a GitHub repository. 🔗 Check it out here: https://lnkd.in/gFjD_dB2 This project sharpened my skills in Python, Pandas, Seaborn, Matplotlib, and EDA, while also showing how data analysis can drive smarter buying decisions. Thanks to my trainer Shankargouda Tegginmani and mentor Abhishek B for their valuable guidance #DataAnalysis #EDA #Python #ElectricVehicles #GitHubProjects #DataDriven

10

Shipping optimization inside an LLM flow with a hard SLA is tough: NP-hard problems, spiky instance sizes, and latency tails mean you can’t rely on “prove optimality.” The fix is to architect for fast feasibility: precompute/warm-start candidates, trim the live model (feature-based fixing/rolling horizon), have the LLM pick a solver recipe and generate a construction heuristic, then run an anytime solve with a firm time limit and return the best incumbent plus gap.

13

n8n vs OpenClaw: Choosing the Right Tool for Production-Grade AI Agents With the OpenClaw hype, a question keeps coming up: “Should we use OpenClaw or n8n for production agent workflows?” The answer isn’t about which tool is “better”. It’s about what kind of control you need in production. n8n works best when: -- Workflows must be deterministic and auditable -- You need clear retries, failures, and rollbacks -- Compliance, logging, and predictability matter -- Humans must stay in the loop OpenClaw shines when: -- Tasks are exploratory or loosely defined -- The agent needs to reason, plan, and adapt -- You’re prototyping or experimenting with autonomy -- Failure is acceptable and reversible In serious systems, n8n (or similar tools) often become the guardrails, while agentic frameworks operate inside controlled boundaries. The future isn’t n8n vs OpenClaw. It’s deterministic orchestration + constrained autonomy.

3

1 Comment

Too Much Thinking Can Break LLMs: Inverse Scaling in Test-Time Compute Recent advances in large language models (LLMs) have encouraged the idea that letting models “think longer” during inference usually improves their accuracy and robustness. Practices like chain-of-thought prompting, step-by-step explanations, and increasing “test-time compute” are now standard techniques in the field. However, the Anthropic-led study “Inverse Scaling in Test-Time Compute” delivers a compelling counterpoint: in many cases, longer reasoning traces can actively harm performance, not just make inference slower or more costly. The paper evaluates leading LLMs—including Anthropic Claude, OpenAI o-series, and several open-weight models—on custom benchmarks designed to induce overthinking. The results reveal a rich landscape of failure modes that are model-specific and challenge current assumptions about scale and reasoning. Full Analysis: https://lnkd.in/gQdf4SE3 Paper: https://lnkd.in/gytEPFAK Project: https://lnkd.in/gtbPebjf Code: https://lnkd.in/g-Buqk2H Video Analysis: https://lnkd.in/g6jGpQBC Anthropic

36

1 Comment

We are strictly back to Feature Engineering. There was a brief golden age where we threw raw data at Convolutional Neural Networks and they just *worked*. We thought we had solved the representation problem. The dream was end-to-end learning. But look at your actual workflow with LLMs right now: - Structuring data for the prompt? That’s just manual Encoding. - Summarizing big projects to fit the window? That’s Dimensionality Reduction. - Chain-of-thought prompting? That’s Feature Extraction. I find myself spending less time coding and more time manually curating inputs to coax a specific output from a black box. The tools changed, but the bottleneck is exactly the same as it was 20 years ago: Garbage in, Garbage out. #GenerativeAI #Engineering #DataScience #HotTake #AI

1

“Developers don’t want consumer search engines. They want programmable primitives to embed into products. Whoever wins here controls distribution into thousands of apps.”

1

From a finite range of symbols, human language can formulate and comprehend an infinite number of sentences. So, can LLM abstraction be achieved through sophisticated pattern matching alone? Nicolas Figay answers: modelling languages like UML and SysML provide formal, verifiable, transmissible abstraction; semantics is explicit and reasoning is guaranteed by construction, not emergent by statistical accident. LLMs can be useful as long as you don't mistake their capabilities for knowledge or understanding. The ingenuity is in the users. https://lnkd.in/eZSiamCc

5