Jonathan Kershaw

Data stacks are quietly splitting into two camps. One camp is getting bigger, louder, more “enterprise.” The other is getting leaner, closer to the warehouse, and suspiciously calm. The loudest signal in 2025 was the big consolidation around ELT with Fivetran and dbt moving pipelines and transformations under the same roof. From a distance, it looks clean. Efficient. Inevitable. 2026 will show whether it actually stays that way. What we’re seeing on the ground is different. Alongside dbt-style workflows, we’ve been actively using Dataform in real client projects. Not as an experiment. As a default choice in several cases. And here’s the truth: for warehouse-native SQL transformations, the results are not worse. Same idea. Same logic. Same outcomes. Models live in the warehouse. Transformations are written in SQL. Dependencies are explicit. Everything is versioned. The analytics layer gets exactly what it needs — and nothing extra. So the question teams are starting to ask in 2026 isn’t “what’s the standard tool?” It’s: what am I locking myself into? To be clear, Fivetran remains a very strong market standard for ingestion. No debate there. And dbt earned its place by defining analytics engineering as a discipline. But when transformations can live closer to the warehouse — with fewer licenses, fewer dependencies, and fewer points of no return — people notice. What’s emerging right now isn’t a rejection of modern data tooling. It’s a preference for leverage. Less platform gravity. More warehouse ownership. Same results. That feels like a very 2026 direction. What do you think?

79

23 Comments

Using LLMs to understand and port Cobol code and other legacy codebases in government, banking, airlines and health care/insurance to modern maintainable code has got to be a multi-billion-dollar startup idea. Or is it just something the existing consultants will do on their own?? Either way, it's time to let those old Cobol jockeys finally retire. AI can finally let us leave the past behind

19

8 Comments

Perplexity just unveiled Perplexity Computer, and it changes everything. We’ve spent years asking AI for answers; now, we’re going to ask for outcomes. This isn't just another model. It is an orchestration engine that operates a software stack exactly like a human teammate would. Why is this a game-changer? 🔹 Multi-Model Intelligence: It doesn't rely on one "brain." It coordinates 19 specialized models at once—routing reasoning to Opus 4.6, research to Gemini, and long-context recall to ChatGPT 5.2. 🔹 Autonomous Sub-Agents: When you give it a goal, it breaks it into subtasks and creates "sub-agents" to execute them in parallel. One agent drafts while another gathers data. 🔹 Isolated Cloud Sandbox: Every task runs in a secure cloud environment with its own filesystem and browser. It interacts with the real web and your tools (GitHub, Salesforce, Notion) without touching your local device. 🔹 Persistent Memory: It remembers your files, past work, and preferences. It doesn't start over; it builds on what it already knows about you. As engineers and leaders, our role is shifting from "Prompt Engineering" to "Outcome Architecting." The computer is no longer the tool we use; the AI is the computer. Okay. How it is different from Multi-agent solution we are doing? In the customer multi-agent solution, you create agents and define the role and responsibilities and agent communication flow. In Perplexity computer, It uses a "Planner-Executor" model. You give it an outcome (not a process), and it autonomously decides how many sub-agents to spawn, which models to assign, and how to handle roadblocks without you pre-coding the "if-then" logic. https://lnkd.in/ga32dNte #PerplexityComputer #AIAgents #FutureOfWork #DigitalWorker #GenAI #TechInnovation #SoftwareEngineering #AI

19

1 Comment

𝗧𝗵𝗲 𝗦𝗶𝗹𝗲𝗻𝘁 𝗩𝗲𝗹𝗼𝗰𝗶𝘁𝘆 𝗞𝗶𝗹𝗹𝗲𝗿 𝗶𝗻 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗧𝗲𝗮𝗺𝘀: 𝗡𝗲𝗴𝗮𝘁𝗶𝘃𝗶𝘁𝘆 𝗕𝗶𝗮𝘀 𝗶𝗻 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗟𝗲𝗮𝗱𝗲𝗿𝘀𝗵𝗶𝗽 Engineers are paid to find what’s broken. That instinct builds reliable systems. But unchecked, it can quietly slow teams to a crawl. From code reviews to RCAs, we’re trained to spot edge cases, vulnerabilities, and failure modes. That skepticism-first mindset is a feature in engineering, but it becomes a liability in leadership. That’s negativity bias at work: our brains overweight negative signals and underweight positive ones. How it shows up in real teams?  • The “fix-it only” feedback loop: When a deploy goes smoothly, silence.  • Risk aversion disguised as rigor: The pain of a potential failure feels far bigger than the upside of improvement.  • Hiring blind spots: We zoom in on red flags and gloss over green ones. Why the 3:1 ratio matters? Research from Dr. Barbara Fredrickson shows high-performing teams need roughly 3 positive interactions for every negative one to truly flourish. That doesn’t mean fake cheerleading. It means being intentional. If feedback only shows up when something breaks, the ratio is already upside down. That’s not rigor; it’s erosion of psychological safety. Ways to counteract negativity bias (without lowering the bar)  • Celebrate non-events: Call out a clean migration, a solid PR, or a quiet week with zero P0s.  • Use pre-mortems, not just post-mortems: Channel skepticism early, during design.  • Balance PR feedback: Encourage reviewers to note what worked, not just what needs fixing. The real takeaway? Engineering excellence isn’t just about reducing failures. It’s about increasing wins. And making them visible. When leaders balance problem-spotting with recognition, teams become more resilient, more creative, and faster over time. So here’s the question worth asking: How do you keep your team sharp without training them to expect punishment for visibility?

5

No logs. No test set. A RAG system returned a superseded regulation for 7 weeks. Nobody knew. It went live on a Monday. Logging and evaluation were both listed as “Phase 2.” Phase 2 never came. The system answered queries. Users trusted it. Filings were made. For 7 weeks. One retrieval path was returning a regulation superseded 11 months earlier. The document store contained both versions. But there was no metadata to distinguish current from archived. The retriever had no way to know. The system had no way to surface it. Nobody noticed. Week 8 — an external audit flagged a regulatory filing. Then came the question: “How many other queries returned wrong answers in the last 7 weeks?” Answer: We don’t know. No logs. No trace. No test set to replay against. It took three weeks to reconstruct what happened. Manual. Painful. Incomplete. The gap wasn’t the model. The gap was two missing foundation disciplines: Observe: Know what your system is doing in production. Eval: Know when it starts doing it wrong. Without those two in place, you are not running a RAG system. You are running a guess with a UI. If you shipped RAG without logging and evaluation, this story is not hypothetical. It’s just scheduled. Observe and Eval sit in the Foundation layer of the RAGBEE™ framework — before hardening, before optimization. Because a system you cannot see and cannot test is not production-ready. Live RAG Architecture Masterclass Saturday · March 21 · 11:00 AM IST Under 60 minutes · No replay Seat confirmation: https://lnkd.in/d5guXJmW #RAGArchitecture #ProductionAI #EnterpriseAI