Richard Demsyn-Jones

Dynamically generated, Ontology-powered fleets of agents. At DevCon 5, Palantir Product Manager Ben Radford walks through the architecture of AI Hivemind — covering the capabilities and tradecraft that enable the creation, orchestration, and optimization of specialized agent teams.

436

12 Comments

With so many young professionals now seeking careers in AI, I'm often asked about how to enter the field. The answer is that there's no one path. At TD and Layer 6 AI, we've built a team of over 200 colleagues whose journeys have led them through different countries, educational backgrounds and working experiences. About 60% of the people we hire are recent university graduates. We look for talent at all levels — bachelor's, master's and PhD — with backgrounds in applied sciences, mathematics, engineering, computer science and more. But sometimes the soft skills are what stands out most. Sumee Seetharaman has done a fantastic job in building the Layer 6 strategy and delivery team with colleagues like Diarra Mbacke who have a drive to learn — and to do it quickly. Read about Sumee and Diarra's journeys to careers in AI here: #AI #innovation #Layer6 #TD https://lnkd.in/gDwVK-ev

155

2 Comments

There’s a major imbalance in 🇨🇦 Canada’s IT job market. While many junior and mid-level engineers are desperately searching for opportunities, companies often prefer hiring more experienced professionals for lower-level roles or short-term contracts under the label of headhunting. To build a healthier ecosystem, employers should focus on nurturing junior and mid-level talent instead of hiring more experienced professionals at lower salaries just to cut costs.

73

The End of Manual Modeling For decades, data modeling was manual by design — a slow, meticulous craft that required expertise and patience. But AI analytics moves too fast for that. Today, data models must evolve in real time. Schema changes, new domains, and user-driven queries arrive faster than humans can reconcile definitions. The solution isn’t to replace modelers; it’s to augment them. AtScale’s AI-powered semantic modeling brings intelligence to the modeling process itself, learning from metadata, lineage, and query patterns to suggest and refine logic at scale. When your business changes daily, your models should too — without breaking trust or governance. Read more from our CTO David P. Mariani as he breaks down how AI is transforming semantic modeling in his latest blog:

1

Unlock the Power of Unified Data with Databricks In today’s data-driven world, organizations need more than just data storage they need data intelligence. That’s where Databricks shines bringing together data engineering, data science, and machine learning on one collaborative platform. Why Databricks Matters: 1. Unified Data Platform: Combines data lakes and warehouses for a “Lakehouse” architecture. 2. Scalable AI/ML: Enables rapid model training and deployment with built-in ML flow. 3. Collaborative Workspace: Data engineers, analysts, and scientists can work together seamlessly. 4. Open & Cloud-Ready: Built on Apache Spark, compatible with major cloud providers (AWS, Azure, GCP). At GrayCell America, we help organizations leverage Databricks to accelerate data transformation, AI adoption, and decision-making. From data silos to smart insights your data journey starts here. #Databricks #DataEngineering #AI #MachineLearning #DataScience #GraycellAmerica #DataAnalytics #DigitalTransformation

11

🚀 In the dynamic tech landscape, Palantir Technologies Inc. (PLTR) and BigBear AI Holdings, Inc. (BBAI) carve out distinct paths in artificial intelligence and data analytics. This comparative analysis dives deep into their strengths, weaknesses, and investment potential. Palantir, with its robust software platforms that support intelligence and defense operations, boasts a market cap of approximately $375B and impressive revenue growth. On the flip side, BigBear AI offers innovative decision-support solutions but grapples with significant operating losses and a much smaller market cap of $2.2B. Key takeaways highlight Palantir's superior profitability and market position, in contrast to BigBear's challenges amid a growth phase. While both stocks show bullish trends, Palantir appears to be the more stable investment choice according to recent analyst opinions. Curious about which company's stock might be the better fit for your portfolio? Explore the full analysis here: https://lnkd.in/eVMPs5ks #Investing #AI #TechStockInsights

1

Databricks didn’t just release another AI feature. They made Custom Agents production-ready. That’s a big deal. Because most organizations aren’t struggling with ideas. They’re struggling with operationalization. The real question isn’t whether AI agents are powerful. It’s whether your data foundation is ready for them. Are you already building on Databricks, or still experimenting outside your core platform? #Databricks #AI #DataEngineering #Lakehouse #EnterpriseAI

4

The insights keep coming at #BigDataCanada2025 and we’re just getting started. 📊 11:00 AM – Monetization Panel Turning data into dollars. In a tough economy, learn how leaders are proving ROI and unlocking new revenue through smart monetization strategies: • Alexander Chan, CFA, MA, Senior Director & Head of Analytics, RBCx • Jinal Patel, Technical Program Manager, TELUS • Ahmed Memon, Senior Enterprise Architect, Co-operators • Bala Gopalakrishnan, Former Chief Data Officer, Pelmorex (The Weather Network) • Saquib Rashid, Staff Customer Engineer, Collibra • Moderated by Neha Balooja, MBA, Sr. Manager, Channel Performance & Marketing Analytics, Rogers Bank 🤖 11:30 AM – Industry Expert Session From Hype to Reality — Josh Good, Vice President, Strategy & Market Development, Qlik, breaks down what it really means to be AI-ready. 🧠 12:00 PM – Case Study What does it mean to preserve legacy in a digital world? David Parker, Founder, Wishplay explores the emotional side of data and the future of AI-powered storytelling. 👥 After the sessions, swing by the Exhibitor Lounge. Connect with fellow innovators, chat with sponsors, and check out the tools real teams are using to get results. You might just find your next collaboration waiting there.

15

Maximally federated or centralized? Address platform governance now or later? These are great questions that we explore with the best training partner in the Palantir ecosystem. CodeStrap's "Code and Connor" Episode 9 releases on August 28, 2025 with our friend Taylor Gregoire-Wright, founder of Ontologize. Ontologize builds teams of Foundry experts at Palantir customers & partners. Our ninth episode focuses on: → Decentralized vs. centralized approaches to technology in enterprise → Journey to “owning the Palantir stack” for customers → Maximizing resource value, simplifying the platform → Aligning incentives to drive innovation → The role of the Palantir champion → Objections to Palantir and how to engage developers → Don’t lift and shift, rather extend the value of IT investments → Ontologize's journey and roadmap → Foundry Marketplace, its fundamentals, and its massive potential value → DevCorner: Public registration system feeding Foundry Be sure to follow us to get the latest information, and schedule a meeting with us through www.codestrap.com!

7

Shining a spotlight on Machine Learning Talent in Canada. Here are some members who recently joined our ML Community: 🍁 William C., Senior MLOps Engineer at Unitary, where he builds and optimizes virtual agent workflows that combine LLMs, custom models, and human oversight to enhance scalability and operational efficiency. 🍁 Yousef Nademi, PhD, Senior Machine Learning Developer at AltaML, where he builds and deploys advanced agentic AI systems, generative models, and MLOps pipelines that drive automation and innovation in real-world data applications. 🍁 Egina Malaj, Ph.D., Senior Manager of Data & AI CoE at Avanade, where she leads enterprise AI strategy and the development of scalable data-driven systems that connect business transformation with applied machine learning innovation. 🍁 Michal Malyska is the Associate Director of MLOps at Fable Therapeutics, where he leads the design and scaling of machine learning infrastructure, along with optimization for model training and inference, to power AI-driven therapeutic research for drug R&D. 🍁 Mitchell McCaig is a Technical Product Advisor at Amii (Alberta Machine Intelligence Institute), where he guides startups in leveraging AI/ML, after earlier roles as a machine‑learning educator, data intelligence analyst and applied‑vision. 🍁 Nafisah Nubah is a Dalhousie University CS student involved in ML research and software development, with experience spanning accessibility tech, teaching, and academic community engagement."

18

1 Comment

Following Andrej Karpathy's viral analysis of AI exposure across every occupation in the US economy, I have been doing the same for Canada this afternoon. Using real data from Statistics Canada and the ESDC Job Bank (NOC 2021), I scored all 516 Canadian occupations on Digital AI Exposure. How much will AI reshape each job in Canada? Building the treemap visualizer now. Will share later today! Stay tuned. 🍁

14

1 Comment

What an incredible week at KDD 2025 in Toronto! The energy, innovation, and depth of discussion were truly inspiring — and it was great to see some old friends and meet new researchers. Here are my key takeaways: 🔹 LLMs & Generative AI — Now With Purpose The focus has shifted from simply building GenAI apps to tackling deployment, efficiency, security, and ethics head-on. 🔹 Data Quality Takes Center Stage The new Datasets & Benchmarks track highlights the importance of high-quality, accessible, and ethically developed datasets for reproducible and trustworthy AI. 🔹 Graph Learning & Time Series Make a Comeback Exciting advances in GNNs and sophisticated time series techniques—often fused with GenAI (I’ll share more at my AIAI talk later this year). 🔹 Applied Data Science is Redefining the Field This year’s Applied Data Science track delivered groundbreaking research with real-world, industry-scale impact. 📚 Interesting Tutorials I Attended/Reviewed Relational Deep Learning (Stanford) — https://lnkd.in/e3q_5ies Training Industry-Scale GNNs with GiGL (Snap) — https://lnkd.in/ecmSGtTY Unifying LLMs & Knowledge Graphs for Medicine & Healthcare — https://lnkd.in/e5dYAknH Graph Foundation Models — https://lnkd.in/epwhBRaS Graph Prompting for Graph Learning Models — https://lnkd.in/ef2NTFfn 🏆 KDD 2025 Awards (Links to Papers) Best Paper – Research: Improving Group Robustness on Spurious Correlation via Evidential Alignment — https://lnkd.in/eFBkhqk2 Runner-Up – Research: Monitoring Robustness and Individual Fairness — https://lnkd.in/eXdetQTA Honorable Mention – Research: SIGEM – A Simple yet Effective Similarity-based Graph Embedding Method — https://lnkd.in/eSgmgAsP Best Student Paper – Research: Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity — https://lnkd.in/eGfHc_ra Best Paper – Applied Data Science: Evaluating Decision Rules Across Many Weak Experiments — https://lnkd.in/e_NgXXim Honorable Mention – Applied DS: Web Scale Graph Mining for Cyber Threat Intelligence — https://lnkd.in/ecys5-ye Best Paper – Datasets & Benchmarks: When Heterophily Meets Heterogeneity — https://lnkd.in/eict3wnA Audience Award (Aug): LLMLight – Large Language Models as Traffic Signal Control Agents — https://lnkd.in/e_Zw29W6 Audience Award (Feb): Theory Instead of Experiment (TIE) – A Creator Valuation System at Tencent — https://lnkd.in/eK45HHZe #KDD2025 #DataScience #AI #LLMs #GenerativeAI #ResponsibleAI #MachineLearning #GraphML #FidelityAssociate

25

Just returned from ICML 2025 in Vancouver, one of the world’s premier machine learning conferences, where I had the privilege of presenting our work on Differentially Private Boxplots. Our core message: Privacy is not just about the algorithms, but about what you can see and interpret. https://lnkd.in/eaif-3b7 It was an inspiring and action-packed week, offering not only groundbreaking research but also a glimpse into the present state and future direction of the field. Some key takeaways from the conference: - AI systems often develop internal models of the world that differ from human expectations. Understanding this mismatch is essential for improving reliability. - The legal landscape around generative AI is still in flux, especially regarding data usage and copyright. Clear frameworks are urgently needed. - As AI becomes more integrated into society, alignment with human values must account for the fact that those values are constantly evolving. - Choosing what AI should optimize for is as critical as how it optimizes. Objective design is foundational across all applications. - End-to-end machine learning systems are now enabling closed-loop optimization and accelerating discovery in science and engineering. #ICML2025

26

Using Databricks is no longer just an option—it's a necessity for enterprises aiming to change data into actionable AI insights quickly and reliably. With Databricks' unified analytics platform, companies can accelerate data processing, streamline AI model development, and deploy solutions faster than ever. But the real game-changer? Access to certified experts who know how to tap its full potential without delays. At KData Inc., we've seen firsthand how rapid deployment of Databricks-certified professionals within 48 hours helps organizations overcome critical skill shortages and meet aggressive project timelines. For example, a major financial services client recently reduced their AI model deployment time by 30% by partnering with our experts. What this means for your business is clear: accelerated advancement, minimized risk, and faster time-to-value from your data initiatives. Whether migrating legacy systems or building AI-driven products, having the right talent and proven accelerators on your side is crucial. Are you ready to change your data strategy with speed and certainty? Let's discuss how using Databricks with certified expertise can deliver measurable impact for your organization.

4

CANSIM (aka 'NDM') is Canada's socio-economic data repository. Basically FRED + BLS + Census Bureau and everything else assembled into a gigantic data cube. There are something like 60k+ tables. StatCan covers a handful at a time in The Daily, but it's impossible to document every update and release across all these series. But what if the marginal cost of coverage was (nearly) zero? The D-AI-LY is an autonomous harness that trawls CANSIM, checking for recently updated or otherwise neglected data series and generates statistical bulletins with visualizations, metadata, and reproducible code (in English and French). As with everyone's holiday projects, it was built with Claude Code, but specifically I wanted to learn how to build custom `SKILL md` files and how to be more deliberate and thoughtful with data provenance and tracking to reduce data errors and hallucination. There are three detailed custom skills that guide data retrieval, generation, and publishing. Skills are an improvement on prompting because they allow you to essentially deeply fine-tune instructions with guided examples without needing to explicitly put that detail into a prompt -- and because of their progressive disclosure, they save context space until the skill is directly invoked by Claude. This allows you to get very specific in your instructions, which is crucial for a system that generates hundreds of articles in a consistent and deliberate visual and narrative style. The most interesting challenge was overcoming persistent and insidious data errors and hallucinations. My approach was to simultaneously layer: 1. SKILL files to carefully fine tune instructions, expectations, skepticism and reasoning 2. Giving CC access to and forcing reliance on specialized tools like the cansim R package 3. Over-engineered data provenance tracking to trace each data point While obviously a toy system, overall I'm pretty happy with the result and it was easy to generate hundreds of these bulletins at scale. Moreover, it's clear to me that SKILLS are the next big thing everyone will get hyped over. SKILLS allow for externalization and commodification of nuanced expertise and I assume we'll start to see some kind of marketplace economy develop around them soon. I wrote up a post describing the SKILLS setup in a lot more detail and another one about the challenge of fighting fake data. Links in comment below. Please let me know what you think!

43

3 Comments