urn,type,status,published_at_utc,post_url,title,excerpt,text,image_url,impressions,reactions,comments,reposts,stats_captured_at 7487164575059058688,post,PUBLISHED,2026-07-26T15:18:59.485Z,https://www.linkedin.com/feed/update/urn:li:activity:7487164575059058688/,Pairwise Testing deserves more attention—especially now,The applied mathematics behind it offers a practical answer to a persistent engineering problem: How do we evaluate a complex system systematically when exhaustive testing is too…,"Pairwise Testing deserves more attention—especially now. The applied mathematics behind it offers a practical answer to a persistent engineering problem: How do we evaluate a complex system systematically when exhaustive testing is too expensive and time-consuming? The key concept is a Covering Array—a mathematical construct that transforms thousands of possible combinations into a smaller, systematic set of test cases while ensuring that every pair of input values appears at least once. In other words: • Combinatorial Science defines the challenge. • Covering Arrays provide the mathematical foundation. • Pairwise Testing turns it into an engineering practice. I previously explored this approach in detail, with practical examples, in my article on applying Combinatorial Science to enterprise data platforms. The concept is even more relevant in the era of Agentic AI and AI Evals. An agentic system may vary across models, prompts, tools, datasets, workflows, memory configurations, policies, and runtime conditions. Testing every possible combination quickly becomes impractical. Evals are also not fully automated. Humans still steer: • Evaluation scenarios • Judgments and expected outcomes • Rubrics and scoring criteria • Risk thresholds • Edge cases and failure analysis Applying combinatorial techniques can help teams design a smaller but more systematic evaluation matrix—reducing model calls, human review effort, cost, and execution time while preserving meaningful interaction coverage. ""Every pair, at least once. That simple principle may become an important foundation for scalable AI evaluation engineering. "" Read the detailed article for the concepts, examples, and enterprise use cases. Article link: https://lnkd.in/g49nJ3wQ Article title ""Applying combinatorial science and discrete mathematics to data-platform engineering and testing"" #AppliedMathematics #CombinatorialScience #CoveringArrays #PairwiseTesting #AIEvals #AgenticAI #AIEngineering #DataEngineering",https://media.licdn.com/dms/image/v2/D5622AQH3GqfLeHMRog/feedshare-shrink_800/B56Z.fAgbCHYAc-/0/1785079138251?e=1786579200&v=beta&t=Y2-gN25g6tgPTsti6Oe-03FvTtuY_7joFy9mPhcGztQ,233,0,0,0,2026-08-02 7487142880839565312,post,PUBLISHED,2026-07-26T13:52:47.180Z,https://www.linkedin.com/feed/update/urn:li:activity:7487142880839565312/,Meet DataGov — my new open-source data governance CLI,"Written in Rust. Built for agents. The goal is simple: one binary, with no Python, Java, Node.js, or Docker dependency, that both engineers and AI agents can run reliably. DataGov…","Meet DataGov — my new open-source data governance CLI. Written in Rust. Built for agents. The goal is simple: one binary, with no Python, Java, Node.js, or Docker dependency, that both engineers and AI agents can run reliably. DataGov is being designed to: - Inspect CSV and Parquet datasets - Profile columns - Parse, format, and transpile SQL across dialects - Scan for PII - Run data quality checks - Generate a normalized JSON governance report Here are some commands you can use: - datagov inspect customers.parquet - datagov profile customers.csv - datagov sql transpile query.sql --from spark --to duckdb - datagov pii scan customers.parquet - datagov report customers.parquet --profile --pii --output report.json Why am I building it? Data governance has been my primary focus for years. As agents increasingly handle data engineering tasks, they require governance tools that provide JSON output, stable exit codes, deterministic behavior, and no interactive prompts. I couldn't find a CLI that merged these agent-ready features with effective data governance capabilities, which led me to create this tool. The guiding principles are: - Deterministic first - Local first - Privacy first PII detection will rely on rules and parsers instead of LLM guesses, ensuring sensitive values remain masked by default and that data stays on the local machine. This is just the beginning; it's an early-stage kickoff rather than a finished product announcement. Bolt 1, the CLI foundation and governance report envelope, is complete, and Bolt 2, focused on dataset inspection, is currently underway. The first milestone will encompass inspect, profile, SQL, PII scan, and report functionalities. Check out the repository at https://lnkd.in/gsuU48M5 #AdaptiveDataGovOps #DataGovernance #OpenSource #Rust #AIAgents",https://media.licdn.com/dms/image/v2/D5622AQFb_gqoYru3nA/feedshare-shrink_800/B56Z.esxnFK4Ac-/0/1785073965877?e=1786579200&v=beta&t=DWkYtea5UcLW0OsG1Ch5b8qpcrV9zO2fMZahTiDBUEM,1134,0,0,0,2026-08-02 7486083742336417792,post,PUBLISHED,2026-07-23T15:44:08.883Z,https://www.linkedin.com/feed/update/urn:li:activity:7486083742336417792/,Introducing the latest addition to my agent fleet,"the youtube-transcriber, now number nine. What it does: Simply provide a YouTube link, and it returns the full spoken transcript of the audio, not just the captions. This means…","Introducing the latest addition to my agent fleet: the youtube-transcriber, now number nine. What it does: Simply provide a YouTube link, and it returns the full spoken transcript of the audio, not just the captions. This means you get the actual audio transcribed from start to finish. Why I built it: In conference talks, important figures are often mentioned in passing. Speakers may only mention a statistic once, and then it's lost. Captions can be auto-generated and may miss crucial information or be altered by marketing teams. I wanted a way to make those numbers searchable, with the speaker and timestamp included. What powers it: - LangGraph with five nodes: resolve → fetch audio → transcribe → write outputs, featuring a cached-audio branch for repeat runs to skip downloads. - faster-whisper (distil-large-v3) operates on my machine, not through an API. - yt-dlp for audio extraction and ffmpeg for normalization. - Pydantic for state management, tenacity for retries, and optional OpenTelemetry (I am using Arize and LangSmith) Each run produces four files: a readable transcript with timestamps, a full JSON file with per-segment timings, an SRT file, and a metrics file. In the first four talks (The Youtube videos, I used for my testing), I processed 67 minutes of audio, generating 10,658 words in about 14 minutes of compute time—roughly 4.5 to 4.9 times faster than real-time, with zero tokens and zero costs, as it operates entirely on my mac-mini. The other eight agents include: job-scout, job-matcher, job-pilot, api-test-generator, diagram-generator, github-pr-digest, privacy-classifier, and linkedin-cover-generator. #AIAgents #LangGraph #Whisper #LocalFirst #AINative",https://media.licdn.com/dms/image/v2/D5622AQFqKGhrGbCNmA/feedshare-shrink_800/B56Z.Ppfn.GUAc-/0/1784821447573?e=1786579200&v=beta&t=tifTNUTUsiD0ri8VB8kEFLsgFnQ2NjYXa6nj3R4ssh0,434,0,1,0,2026-08-02 7486074529379913730,post,PUBLISHED,2026-07-23T15:07:32.343Z,https://www.linkedin.com/feed/update/urn:li:activity:7486074529379913730/,"The landscape of agent frameworks is rapidly evolving, with new names appearing…","every few weeks: OpenAI Agents SDK, LangGraph, CrewAI, PydanticAI, LangChain, Claude agents, and more. Despite the varying terminologies and abstractions, the underlying execution…","The landscape of agent frameworks is rapidly evolving, with new names appearing every few weeks: OpenAI Agents SDK, LangGraph, CrewAI, PydanticAI, LangChain, Claude agents, and more. Despite the varying terminologies and abstractions, the underlying execution pattern remains largely consistent. Currently, there is no single industry-standard agent lifecycle, but a common pattern is emerging that is more beneficial to learn than memorizing the vocabulary of any one framework. Here are the eleven stages, typically in the order they run: 1. User request 2. Context assembly 3. Memory retrieval 4. Knowledge retrieval 5. Planning 6. Tool or skill selection 7. Execution 8. Response generation 9. Evaluation 10. Reflection 11. Optional memory update It's important to distinguish between evaluation and reflection. Evaluation assesses whether the output met the task requirements, while reflection considers the process and what changes might be needed for future runs (Feedback loop and Agent Training). Additionally, knowledge and memory are distinct; knowledge is derived from documents, databases, and APIs, whereas memory stems from interactions, preferences, and past decisions. Context is built from both for the current request only. Response generation, which is stage 8 of 11, is the visible part of the process, but it is often mistaken for the entirety of the system. I have compiled the full lifecycle into a one-page cheatsheet, as shown in the pdf document below. #AIAgents #AgentEngineering #GenAI #RAG #AIEngineering …more Your document has finished loading
This single line is far more valuable than storing the entire transcript. Thus, the loop becomes: reflect, distill, ask for confirmation, store as structured memory, and retrieve it next time. No fine-tuning is needed in that process. There are essentially two loops at play, and conflating them is a mistake. The first is user-level, capturing my preferences—writing style, palette, level of detail, layout—enhancing the agent's performance for me. The second is product-level; if numerous users request the cover generator to reduce clutter, that indicates a need for a new version of the core skill itself. Memory personalizes the agent, while reflection enhances the system. Repeated feedback evolves the skill. Building the agent is only half the job; teaching it how to improve is the other half. Agent Engineering is gradually evolving into Memory Engineering, as memory is not merely the full transcript but the distilled knowledge that deserves to persist beyond the session. #AgentEngineering #AIAgents #MemoryEngineering #GenAI #AIEngineering",https://media.licdn.com/dms/image/v2/D5622AQFQfY_qSk4T2w/feedshare-shrink_800/B56Z.O5WuzKwAc-/0/1784808828059?e=1786579200&v=beta&t=lidb2vLHm7TCHbqir2NHK1eh5tsJ8EyYyRmsaXtCCsU,599,1,0,0,2026-08-02 7485498869477310464,post,PUBLISHED,2026-07-22T01:00:04.327Z,https://www.linkedin.com/feed/update/urn:li:activity:7485498869477310464/,"Every week, new ideas emerge around AI-SDLC, spec-driven development, agentic…","engineering, AI evaluations, governance, observability, and AI-native delivery platforms. The engineering side is gradually converging around common practices, methodologies, and…","Every week, new ideas emerge around AI-SDLC, spec-driven development, agentic engineering, AI evaluations, governance, observability, and AI-native delivery platforms. The engineering side is gradually converging around common practices, methodologies, and supporting ecosystems. However, AI is also transforming the economics of software delivery. If AI significantly alters delivery speed, team structures, engineering effort, and quality, how should projects be estimated, priced, and valued? This question extends beyond just business or sales; it is crucial for architects, engineering leaders, pre-sales teams, delivery managers, and consulting organizations. The estimate prepared during pre-sales ultimately becomes the delivery commitment that engineering teams must fulfill. It appears that while the engineering conversation is becoming increasingly structured, estimation, pricing, value realization, and commercial models are evolving more quietly. Are our commercial practices converging as swiftly as our AI engineering practices? I am interested in how others are approaching this transition.",https://media.licdn.com/dms/image/v2/D5622AQHVnn-cTZsXUA/feedshare-shrink_800/B56Z.HVjqPGQAg-/0/1784682003337?e=1786579200&v=beta&t=SkxyHvkfhZAiwHtef1w3X_eeojOhZxKKUQfW5STND9s,566,0,1,1,2026-08-02 7484656081412001792,post,PUBLISHED,2026-07-19T17:11:07.994Z,https://www.linkedin.com/feed/update/urn:li:activity:7484656081412001792/,I am setting up a production line for my open-source work — and the workers on…,"it are AI agents. Every project needs a website, docs, a blog, and an API reference. I did not want to hand-build that each time, and I do not want my agents to either.…","I am setting up a production line for my open-source work — and the workers on it are AI agents. Every project needs a website, docs, a blog, and an API reference. I did not want to hand-build that each time, and I do not want my agents to either. TemplrPress is the publishing station on that line. It is my opinionated, open-source content publishing system — I built it for my own needs first, and now I am giving it to the community. A 15 MB Go binary: markdown in, product website out. It started in 2021 as Zypress, built on Vue.js/Nuxt.js CMS. Five years of gradual work later, agents can operate it end to end: - Docs and OpenAPI reference in one place, free. - PDF export in pure Go — no headless browser. - A JSON API serving the same content. - Paste a link in WhatsApp or Slack and it shows a proper preview card. - /llms.txt plus a committed code map, so agents read the site and the repo without scanning. Fully open source, MIT license. Built on Golang and NextJS/TypeScript — the UI compiles into the binary, so the server needs no Node, no database, no build step. One config file, one station done. More stations coming. Code: https://lnkd.in/eZxzGfY2 Demo: https://lnkd.in/e2iYASgB #GoLang #OpenSource #AIAgents #DeveloperTools #LLM",https://media.licdn.com/dms/image/v2/D4E22AQE7PjQfjdlATw/feedshare-shrink_800/B4EZ97XC4ZJMAg-/0/1784481066864?e=1786579200&v=beta&t=a61LCZURmSdB7Mucjsp-Y7l9ay55eua0_ey7hTLJwCY,290,0,0,0,2026-08-02 7483510885441937408,post,PUBLISHED,2026-07-16T13:20:31.998Z,https://www.linkedin.com/feed/update/urn:li:activity:7483510885441937408/,"Recently, while exploring modern AI evaluation frameworks, I had a realization","Initially, the categories seemed like a collection of technical metrics—valuable for testing model behavior, comparing outputs, and enhancing performance. However, the deeper I…","Recently, while exploring modern AI evaluation frameworks, I had a realization. Initially, the categories seemed like a collection of technical metrics—valuable for testing model behavior, comparing outputs, and enhancing performance. However, the deeper I delved, the more I understood that they represent something much more significant: the dimensions of AI trust. - Security: Can the system protect sensitive information and resist misuse? - Safety: Can it reduce harmful, toxic, or biased outcomes? - Quality: Is the response correct, relevant, grounded, and reliable? - Conversation: Did the system understand the user and help accomplish the intended task? - Trajectory: Did the agent choose the right tools, follow the right process, and reach the outcome in a defensible way? These are not just evaluation metrics; they are measurable signals indicating whether an enterprise should trust an AI system in production. This perspective shifts how we should approach AI engineering. Data Governance defines what trustworthy data should look like. AI Engineering builds systems that utilize that data. AI Evaluations provide evidence that those systems behave as intended. Governance without evaluation remains a set of expectations, while evaluation without governance lacks business and regulatory context. AI engineering without either may yield an impressive system, but not necessarily one that an organization can deploy responsibly. This forms the basis of my thoughts on Adaptive Data Governance: governance, data quality, engineering, risk, and continuous evaluation should not function in isolation. They must create a connected feedback loop that evolves with the data, business, regulations, and the AI system itself. Trust cannot simply be declared at the end of development; it must be designed, measured, documented, and continuously earned. Trustworthy AI is not a feature of the model; it is an engineering outcome.",https://media.licdn.com/dms/image/v2/D4E22AQGYcPxLbSx8dw/feedshare-shrink_800/B4EZ9rFfpFK4Ac-/0/1784208030549?e=1786579200&v=beta&t=C3NLJ3Qw5IRKN5oyTjQ8TrB2dcF3SxPC5QOPUTTj59w,300,0,0,0,2026-08-02 7471004221698543617,post,PUBLISHED,2026-06-12T01:03:30.999Z,https://www.linkedin.com/feed/update/urn:li:activity:7471004221698543617/,Why pay for frontier models when NVIDIA hosts Nemotron 3 Ultra for free,I plugged it into GitHub Copilot CLI in under a minute — swipe through the cheatsheet to set it up yourself. 👇 #AI #GitHubCopilot #NVIDIA #Nemotron #DeveloperTools …more Your…,"Why pay for frontier models when NVIDIA hosts Nemotron 3 Ultra for free? I plugged it into GitHub Copilot CLI in under a minute — swipe through the cheatsheet to set it up yourself. 👇 #AI #GitHubCopilot #NVIDIA #Nemotron #DeveloperTools …more Your document has finished loading