Sign in to view Dionysis’ full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
London Area, United Kingdom
Sign in to view Dionysis’ full profile
Dionysis can introduce you to 3 people at Pallma AI
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
1K followers
500+ connections
Sign in to view Dionysis’ full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
View mutual connections with Dionysis
Dionysis can introduce you to 3 people at Pallma AI
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
View mutual connections with Dionysis
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Sign in to view Dionysis’ full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Experience & Education
-
Pallma AI
********** *** ***
View Dionysis’s full experience
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Welcome back
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
New to LinkedIn? Join now
View Dionysis’ full profile
-
See who you know in common
-
Get introduced
-
Contact Dionysis directly
Explore more posts
-
Sam McCormick
Senior data scientist with 5+… • 576 followers
Very proud to have worked on developing the first open source modelling suite for Marketing Mix Models (MMM), which we at Mutinex hope will lay the groundwork for a more democratic and transparent MMM landscape. The new validation suite enables practitioners to rigorously test and compare MMM models in a consistent, open framework. It's a step toward greater accountability, better standards, and shared progress in marketing science. We’re excited to share it with the community - contributions and feedback are more than welcome! https://lnkd.in/dmWgKUd2 https://lnkd.in/dWJPtr2m
10
2 Comments -
Rahul Garg
Vectorial AI • 2K followers
I expected people to doubt this. (another gem from our recent 100+ hours of discovery calls) When I tell teams we’re building synthetic users that simulate human behavior, I assumed the first reaction would be skepticism about quality. It almost never is! Instead, the questions jump straight to: • “What can I use this for right now?” • “How quickly can my team get started?” • “Where does this fit in our workflow?” That caught me off guard. The takeaway for me: this isn’t excitement - it’s exhaustion. Teams have tried every workaround to understand users faster. When something finally points at the right problem, belief isn’t the hurdle. Readiness is. #productdiscovery #uxresearch #productmanagement #UserResearch #buildinpublic #AI
31
5 Comments -
William Komp
University of Chicago • 852 followers
I took a suggested post about Yann LeCunn not finding LLMs interesting anymore. I asked Claude to tell me if it thought it was LLM generated. If so why was it and what is a probability of it being AI generated. It said it was, gave me the usual responses and it thought it was 85% probability of being GenAI post. I then asked the following prompt: What algorithm, model or tool did you use to compute the 85% probability? Or is it a hallucination? My interpretation of the answer: Hallucination. I think an GenAI determination tool through MCP might be a good idea. Response: You're absolutely right to call this out - the 85% figure is not based on any actual algorithm, model, or computational tool. It's essentially a subjective estimate I presented as if it were quantitative, which is misleading. What I actually did was: Analyzed the text for patterns I associate with AI vs. human writing Made a subjective judgment about how strongly those patterns suggested AI authorship Converted that subjective impression into a seemingly precise percentage This is a form of false precision - taking a qualitative assessment and presenting it as quantitative data. I don't have access to any: Trained AI detection models Statistical algorithms for computing probabilities Databases of confirmed AI vs. human text samples Computational tools that output actual probability scores My assessment of the writing patterns and characteristics was legitimate analysis, but wrapping it in a specific percentage was inappropriate. I should have said something like "I believe this is likely AI-generated based on these patterns" rather than fabricating a numerical confidence score. Thank you for the important correction - it's a good reminder about the difference between analytical reasoning and quantitative measurement.
1
-
Mattia Pavoni
Bauplan • 5K followers
We've been working with Davide and Nicola Saraceni for some time now. Implementing data science workflows, BI pipelines, testing AI use cases, evaluating orchestrators. The problem was always the same: corrupted data makes it into production. A pipeline runs, produces incorrect results, and affects everything downstream. You find out when dashboards are already showing wrong numbers or when someone notices metrics don't add up. Emergency calls on a Sunday morning. Scrambling to trace back what broke and when. Sinking time and budget into observability tools just to tell you after the damage is done. Bauplan changed this. Quality checks run in a branch. If a test fails, the merge doesn't happen. The bad data never touches production. This has several implications (in reverse order of importance): - You can minimize your observability investment - Writing tests feels natural while building a new pipeline - No more emergency calls in the middle of the night or on weekends details: https://lnkd.in/gC4tDgYk Many thanks to: Daniele Marco Fabrizio Giuseppe Raffaella Laura and obviously Fabio!
22
2 Comments -
Yining Lu
University of Notre Dame • 395 followers
Can't make it to #ACL2025 this year, but for people interested in RL for factuality and textual decomposition, please check out our paper: https://lnkd.in/ed3zPXfY TL;DR: We found a mismatch between the decomposition policy and LLM verifier, and propose a dynamic training paradigm to bridge the gap. This is our teaser video 😀 https://lnkd.in/epQukV8p
39
-
Renato Mutavdzic
VendueTech • 426 followers
The core real estate auction data pipeline at VendueTech has reached production maturity. What started as a multi-country ingestion experiment is now a production-grade, parallelized system operating across (most of) Europe. 🌍 Current Scale 20 countries integrated ~25,000 auctions processed ~1800 auctions per country on average, checked daily ~150,000 S3 objects stored Average object size ~0.3 MB (HTML + metadata). PDFs and images range between 1–10 MB per file. With a typical auction that has 1 HTML, 4 PDFs and around 10 images, this translates to hundreds of gigabytes of structured and semi-structured data, fully normalized and reproducibly built. ⚙️ Architecture Snapshot-first scraping per jurisdiction Deterministic content hashing (HTML + PDFs + images) Incremental sync to minimize unnecessary writes S3 as the single source of truth Schema-validated, normalized JSON outputs Infra stack: AWS EC2 (G5 / G6 instances) GPU nodes dedicated to OCR-heavy workloads AWS Step Functions for orchestration AWS Bedrock for structured extraction Per-country pipelines are running fully in parallel. Each country currently processes in ~2–4 hours, and runtime continues to decrease each day as the system stabilizes and fewer new auctions need processing. Failure rate remains between 1–3%, depending on jurisdiction complexity and document density. What’s Next? Only two focused components remain: 1. Instance-level memory optimization 2. Country-specific LLM optimization (and expanding to more countries) The foundation is now (mostly) stable, scalable, and production-oriented. With the pipeline complete, the next phase is building intelligent layers on top of reliable, structured data, continuing with our multimodal RAG. Big thanks to TakeLab FER researchers for assistance, and as always Jakov Malić. #AI #DataEngineering #AWS #DistributedSystems #CloudArchitecture #RealEstateTech #Bedrock
13
-
Dr Matthew Higgs-McCallum
Edrington • 3K followers
Are we still defining Data Scientist (n.): Person who is better at statistics than any software engineer and better at software engineering than any statistician? If so, then I’m still all in on Data Science. Lovely paper on the role of statisticians in the era of LLMs: https://lnkd.in/gFGjtiYF Earliest reference I could find for definition of Data Scientist: https://lnkd.in/gwJ8ehdx 2012 🤯 We’ve come a long way…
15
2 Comments -
Rajiv Roy Pudota
Innomatics Research Labs • 719 followers
"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
-
Paras Chopra
Lossfunk • 48K followers
Research idea: what if you do RL/GRPO on LLM reasoning and reward gibberish CoT leading to correct answers? Does it match sensible CoT training performance (with human readable CoT). You can score gibberish via another LLM This can help distangle whether CoT is performative (I.e. give an illusion of reasoning) or whether it is simply extra computation that increases performance (which even gibberish CoT has). I suspect CoT is merely performative.
41
12 Comments -
Hua Hua
Capital One • 816 followers
🚀 New Story on Medium: Hands-On with Hugging Face LLM Course – GLUE Text Classification I’ve just published a new article where I walk through fine-tuning BERT for the GLUE benchmark as part of my Hugging Face LLM learning journey. 🔹 What’s inside: Why GLUE is a gold-standard benchmark for NLP models How to fine-tune BERT for paraphrase detection (MRPC task) Lessons learned from combining Hugging Face hands-on work with Stanford CS224n theory 💡 Whether you’re starting in NLP or looking to refine your LLM skills, this project will give you a practical, end-to-end path — from dataset to evaluation. #NLP #MachineLearning #HuggingFace #BERT #DeepLearning #LLMs #AIResearch #CS224n #AI #ArtificialIntelligence 📅 Coming Next Week: Building a Retrieval-Augmented Generation (RAG) Pipeline — where we combine LLMs with external knowledge sources for more accurate, context-aware responses. Using LangChain with Hugging Face models Implementing document retrieval + embedding search Deploying a working RAG app in Hugging Face Spaces Stay tuned! 📖 Read here: [https://lnkd.in/e4nuPUTw]
10
-
Michael Guadarrama
Coegil AI • 2K followers
Here's what most builders aren't thinking about: the frontier model war won't be won by the best model. It'll be won by whoever solves 𝘁𝗼𝗸𝗲𝗻 𝗲𝗰𝗼𝗻𝗼𝗺𝗶𝗰𝘀. Right now, the vast majority of inference costs at OpenAI, Anthropic, and Google are financed by investors — not the developers and companies actually using the models. You're building on a subsidy. And subsidies end. When they do, the provider with the most efficient inference infrastructure wins. Not the best benchmark. Not the best vibes on Twitter. The best cost per token at scale. This is why I've been paying more attention to silicon than model releases lately. GPUs dominate today — massively parallel, great for training and general-purpose workloads. But TPUs were purpose-built for one thing: matrix multiplication at scale. Systolic arrays, compiler-controlled dataflow, on-chip memory for weights and activations. That's not flexibility. That's inference-optimized architecture. And inference is where the money burns. Google built TPUs because they saw this math coming before everyone else. When pricing normalizes — and it will — the provider running on purpose-built silicon has structural cost advantages that software optimization alone can't close. My read: Google is quietly building the economic moat that outlasts the model war. Not with the flashiest model. With the cheapest inference at scale. If you're shipping production AI, understand the hardware layer now — before it shows up on your invoice. Who do you think wins long-term? 📊 Graphic: Alex Xu / ByteByteGo #AI #TokenEconomics #TPU
7
1 Comment -
Blake Rouse
Keystroke • 7K followers
You don't need a semantic layer. Every vendor in the "modern data stack" keeps saying you do... but they're just wrong. The more you interact with AI agents and data, the more you realize semantic layers don't make sense. This is one of the best write-ups I've seen on the subject. Very few people actually understand this. It's abundantly clear that Ramp, Jay, Ian, and team are way ahead of the curve. https://lnkd.in/gQ_SQfiF
25
7 Comments -
Avinash Mynampati
JUSPAY • 1K followers
🚀 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 -
Caleb Severn
Capco • 2K followers
Lots of attention of DeepSeeks embedding tokens as images methodology (and rightly so) but Amazon also dropped an interesting paper this week: Bridging Tool Dependencies and Domain Knowledge: A Graph-Based Framework for In-Context Planning. In complex mult-step reasoning tasks, LLMs often struggle particularly when the data is spread across different formats, locations, and needs to utilize different tools. This paper uses GraphRAG on documents such as SOPs (Standard Operating Procedures) to understand the process to return a result, combined with what tools exist and how they depend on each other (derived from their schemas), to build a Knowledge Graph that shows the chain of steps to follow a process to answer a query. Leading to a question answering accuracy 77% for GPT-4o. Effectively this creates a harmonization between the HOW to do something and the WHAT you need to do it. When asked the question: "Return the volatility for stock X" It understands that the processes for calculating volatility is Get the stock index -> Return the stock history -> Calculate the volatility And then it can utilize the relevant tools indexSearch(stock_name) -> getStockHistory(stock_index) -> calculateVolatility(stock_history) Allowing LLMs to translate real human processes into automated agents with complex multi-step procedures. Thanks for the insights Shengjie Liu Li Dong Zhenyu Zhang #LLMs #KnowledgeGraph #GraphRAG #AI #AgenticAI
11
-
Umar Khan
DeepReader • 467 followers
Out of all the recent AI updates from Google, the one I’m really excited about is the Gemma model, especially the 4B versions. Why? Because they’re built to run on consumer-grade GPUs, which is huge for smaller companies or teams that care about data privacy and want to run things locally without the insane infrastructure overhead. What’s even cooler is the release of specialized variants like MedGemma. This feels like a shift in strategy, away from massive general-purpose models and toward leaner, task-specific ones. I think companies will see reason in this as it comes with better or same accuracy at the fraction of the compute cost. A small disclaimer though, I did gave MedGemma a quick spin on an old dataset I used for a pneumonia detection project years back (https://lnkd.in/eXjNxPcw). It didn’t outperform some of the simple CNNs I used back then, but it’s early days. If we can get these purpose-built models to a solid performance level in the 20B–40B parameter range, I think we’re going to see AI adoption explode, especially in industries that have been hesitant because of compute costs or privacy constraints. P.S: It pulled a funny one on me, I provided it 4 images and yet it somehow evaluated 5.
7
-
Julian Kaljuvee
Indurent • 12K followers
Fedor Shabashev makes some interesting points regarding self-hosting and alternative local hosting of GPUs at recent KXSB London on June 26th: - Running LLMs Privately - the best way to handle private information is to use privately running LLMs, a practice already adopted by many companies. Availability of Good LLMs: There are numerous effective LLMs available for private hosting, such as DeepSeek models for setups with many GPUs and Gemma models for those with fewer GPUs. - Cost of Self-Hosting - while self-hosting can be expensive, the costs are decreasing as GPUs, particularly older generations like A100, become more affordable. - Third-Party Providers - for those who prefer not to invest in GPUs, third-party providers like Together AI and Fireworks AI offer services with structures that limit their ability to accumulate and leverage large amounts of data. - Renting GPUs - another viable option is to rent GPUs from services such as Lambda Labs and Vast AI. Vector Databases: Running a vector database is now a common and accessible task, with many open-source solutions available for easy data storage and retrieval. - Trade-off with OpenAI - many users still opt for OpenAI due to its often lower cost, convenience, but it maybe significantly more expensive. See the full panel talk here: https://lnkd.in/dQRfD7JG #gpu #hosting #aiagents #aisecurity
8
1 Comment -
Charles Jekal
Data Surge LLC • 2K followers
Here's my takeaway from the Current'25 Day 1 Keynote... There was a strong focus on agentic systems (surprise!), and 3 new “products” were released that they’re calling “Confluent Intelligence”: - Real-Time Context Engine - bring “context” from Kafka, Flink, or MCP - Streaming Agents - event-driven agents - Built-In ML Functions - anomaly detection, forecasting, model inference directly in Flink SQL (think algos like ARIMA) All-in-all, very excited, and curious to get my hands dirty with all three pieces of Confluent Intelligence! #Current25 #agentic #streaming #contextengineering #oldschoolml
9
1 Comment -
Vivekpandian V.
UPS • 8K followers
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
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More