Comet Assistant Upgrades: Improved Performance and Permissions

Task-first IA: what is the user trying to do right now? Docs navigation often mirrors products and features. Users do not. They arrive with a job in mind and want the shortest honest path to done. So aim the nav at those jobs first. Name pages with verbs. Give each task a clear starting point and an end-to-end “golden path” that’s runnable now. Re-test that path at every new release so it 𝘴𝘵𝘢𝘺𝘴 runnable. If the supporting concepts, options, or edge cases are truly necessary to document, then understand that they are side trips and link out to them. Don't front-load them. When this works, you can see it in your metrics. Time to first success drops. People stop bouncing from the search or chatbot to the page and back again. Support gets fewer “where do I start?” tickets. And the docs look less like an org chart and more like a toolbox. Features matter. Organize those features in service of the user's work, not your own. #TechnicalWriting #InformationArchitecture #DocsAsCode #DeveloperExperience #ContentStrategy

Weaviate's 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 is one of the most comprehensive guides I’ve seen. Covers: → Agents – strategies, context window challenges → Query Augmentation – rewriting, expansion, and decomposition → Retrieval – how chunking and pre/post-processing impact precision → Prompting Techniques – from classic to tool-aware methods → Memory – principles for agent memory → Tools – evolution from prompts to actions Grab a copy→ https://lnkd.in/eiZ5thfd

I’m excited to share a small project I’ve been working on using agentic AI. This proof of concept demonstrates how an AI agent, built with Google’s Agent Development Kit (ADK) and deployed through Vertex AI Agent Engine can transform the way teams discover data insights. The agent in this PoC connects directly to the provided BigQuery datasets and describes its contents through a chat interface. No SQL required. This achieves a reduction of time-to-insight for routine analytical queries. GitHub Page: https://lnkd.in/ej48GCWw Repository: https://lnkd.in/eZkGwmZk

Building and testing a few agent setups lately, and I’ve noticed something — most people overcomplicate it. They jump into prompts and models before defining what the agent is even supposed to do. This visual from Dr. Maryam Miradi captures the flow pretty well. ➡️ Start with purpose. - If you can’t write down what your agent’s role is, you don’t have a design yet. - When the goal is clear, the rest — input schema, tool use, and memory — starts making sense. ➡️ Keep structure non-negotiable. - Schemas aren’t optional. - Define how data comes in and how it should look when it goes out. - It saves you hours of debugging and keeps responses predictable. ➡️ Reasoning is just controlled access. - An agent isn’t “thinking.” It’s deciding which tool to use for a task — search, calculator, or retriever. - Give it the right tools and boundaries, and it’ll look smart. ➡️ Memory gives it context. - Short memory handles the flow in one session. - Retrieval memory lets it pick up from the past. - Choose based on what matters more — speed or depth. ➡️ Multi-agent setups only help if your task is complex. - If you’re coordinating planners, researchers, or checkers, that’s when frameworks like CrewAI or Swarm make sense. - Otherwise, it’s just extra noise. ➡️ UI isn’t for show. - That simple Streamlit or Gradio app helps you see what’s breaking. - Until you interact with it, you won’t know what’s off. ➡️ Evaluation is where maturity shows. - Logs, benchmarks, feedback loops — they’re boring but essential. - That’s how you keep your system reliable over time.

Over the past few weeks, I’ve been building something I’m genuinely proud of , an Agentic AI-based Post-Surgery Recovery Agent developed entirely from scratch using Python on Replit. This project explores how agentic reasoning can enhance healthcare analytics by enabling systems that learn, adapt, and respond intelligently over time. It combines a UCI-trained Logistic Regression model that helps in classifying patient outcomes with an adaptive feedback system that continuously monitors pain levels, activity, and sleep patterns in real time. The agent learns each user’s recovery baseline through z-score–based deviation analysis, detects subtle changes or recurring patterns, and provides clear, contextual insights supported by visual trend charts. Every component , from data ingestion and model training to analytics, visualization, and reporting , operates entirely within Replit, making it a self-contained, explainable AI system for personalized recovery monitoring. You can explore the complete project and source code here: https://lnkd.in/gwu5AtvR

The tools may be automated, but the support is human. That’s by design 💙 Most compliance tools are hard to integrate, harder to trust, and nearly impossible to get help with when things break. At spektr, we’ve tried to build the opposite: → Integrations that feel native to your stack → A team that shows up when it matters → A setup process that doesn’t drain your team’s time or energy We appreciate every customer who shares their experience. It helps us keep raising the bar! Curious to see what this looks like in practice? Try it out for yourself in a 2-min clickable demo: https://lnkd.in/dHmX-ArP

You've got a software product that's packed with impressive features. It's your baby, and you understand every nook and cranny of it. But here's the catch - your potential customers don't. They're not interested in the nitty-gritty technical details... they just want to know what's in it for them. Now what I'm saying is... you need to translate your software features into benefits. Think about how your product solves a problem or makes the user's life easier. Does it save them time? Does it make a complex task simple? Does it provide peace of mind? That's your hook. For example, instead of saying, "Our software has an advanced AI algorithm," you could say, "Our software uses AI to automate your tasks, freeing up your time to focus on what matters most." See what I mean? Your ad copy isn't about impressing with technical jargon. It's about connecting with your audience's needs and desires. So, next time you're writing ad copy, ask yourself: "So what?" So, your software has feature X... so what? What does that mean for the user? Answer that question, and you've got yourself compelling ad copy. My area of expertise if you need more advice, let's chat on a free consultation: https://lnkd.in/d9-U4Ajg

Our team has spent the last few months rethinking what great support tools should feel like. Today, we’re shipping four new AI features in Plain that bring that vision to life. What I love most about this release is how naturally the pieces work together. Here’s what’s new: 💬 𝗔𝗿𝗶 — 𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗲𝗱 𝗦𝘂𝗽𝗽𝗼𝗿𝘁 𝗬𝗼𝘂 𝗖𝗮𝗻 𝗧𝗿𝘂𝘀𝘁 Answers repetitive questions using your docs and past tickets. No guessing, no hallucinations. 👨💻 𝗟𝗼𝗼𝗸𝘂𝗽 𝘄𝗶𝘁𝗵 𝗖𝘂𝗿𝘀𝗼𝗿 — 𝗖𝗼𝗱𝗲-𝗔𝗰𝗰𝘂𝗿𝗮𝘁𝗲 𝗔𝗻𝘀𝘄𝗲𝗿𝘀 Connects to your codebase so support can answer technical questions without interrupting engineering. 🦸 𝗦𝗶𝗱𝗲𝗸𝗶𝗰𝗸 — 𝗬𝗼𝘂𝗿 𝗧𝗲𝗮𝗺’𝘀 𝗖𝗼-𝗣𝗶𝗹𝗼𝘁 Drafts replies, summarizes threads, and surfaces the right docs — all in your inbox. ✨ 𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀 — 𝗙𝗿𝗼𝗺 𝗖𝗼𝗻𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗼𝗻 𝘁𝗼 𝗥𝗼𝗮𝗱𝗺𝗮𝗽 Analyzes support conversations automatically and highlights patterns your product team can use. This brings us closer to a world where support teams have the clarity, context, and tools they’ve always deserved. A big day for Plain, and a reflection of the thoughtful work the team has put into every part of this release.

Keep your annotations clean from the start with Interactive Validation Reports, which provide annotators with real-time feedback that flags errors, enforces consistency, and prevents faulty data from slipping through, all without disrupting the workflow. Why it matters: ✅ Catch errors before they progress ✅ Enforce consistency across frames & sequences ✅ Save time with location-linked fixes ✅ Keep working without disruption Quality, speed, and compliance; all in one workflow. Learn more: https://lnkd.in/dGK4gdah #InteractiveValidations #3DPointCloud #ValidationReport #Annotation #realtimefeedback

The other day I shared a post mentioning tool sprawl and got some great questions about how to put the genie back in the bottle. Most teams feel it long before they have a name for it. It looks like: → Blind spots or making decisions with only part of the story (because that’s all you can get) → Competing versions of “truth” (because teams rely on different sources) → Extra manual work (because someone is stitching numbers together behind the scenes) That friction often turns into “we need 𝘣𝘦𝘵𝘵𝘦𝘳 tools,” but it's rarely the answer. Another tool without a plan is just another place for information to break down. If you’re feeling the pain, take a minute to zoom out and define how information should actually move through your business: 1. 𝐒𝐭𝐚𝐫𝐭 𝐰𝐢𝐭𝐡 𝐝𝐞𝐜𝐢𝐬𝐢𝐨𝐧𝐬. What are the weekly or monthly decisions that matter most? What information do you need to make them well? Everything else should revolve around this step. 2. 𝐌𝐚𝐩 𝐰𝐡𝐚𝐭 𝐞𝐱𝐢𝐬𝐭𝐬 𝐭𝐨𝐝𝐚𝐲. Where does that information come from, where does it live, and how does it flow? You’re not building a giant diagram. You’re spotting friction, duplication, and tools doing more than they were designed for. 3. 𝐈𝐝𝐞𝐧𝐭𝐢𝐟𝐲 𝐭𝐡𝐞 𝐠𝐚𝐩𝐬. Where does information get lost, delayed, or manually reworked? These gaps are usually why teams spend time reconciling instead of using. 4. 𝐃𝐞𝐜𝐢𝐝𝐞 𝐰𝐡𝐚𝐭 𝐭𝐨 𝐜𝐨𝐧𝐧𝐞𝐜𝐭, 𝐬𝐭𝐫𝐞𝐚𝐦𝐥𝐢𝐧𝐞, 𝐨𝐫 𝐩𝐡𝐚𝐬𝐞 𝐨𝐮𝐭. Start with achievable, high-leverage fixes: a cleaner workflow here, a simple integration there, or retiring the tool that no longer earns its keep. The key is small, focused choices that move you toward the goal: having what you need, when you need it. Bite-sized progress beats a year-long overhaul that no one has time for. Use the same steps when evaluating new tools: ground the choice in the decisions and actions you need to take, understand how the information fits into your existing system, and map out what the new flow will look like.