Hey everyone!Just brushed up on my SQL fundamentals again while sitting through 8-hour-LONG!!!! tutorials. 😭I’ve noticed that even some paid courses don’t cover very basic concepts despite having 6+ hours of content.So, I decided to create a single ".ipynb" file that covers SQL from fundamentals to advanced concepts, with hands-on examples throughout.Here’s the link: https://lnkd.in/gsHP_wx9Let me know what you think And if you ever have doubts while learning, feel free to irritate me with DMs. 😂Hope it helps someone!

You can now run a 2.8 TRILLION parameter model on a 4GB GPU.It’s called AirLLM, an open-source tool that uses "Layer-wise Inference." It only loads one layer onto your GPU at a time. so the VRAM you need depends on the layer size, not the model size.No quantization. No distillation. No pruning.→ DeepSeek-V3 (671B) on 12GB→ Llama 3.1 405B on 8GB→ Kimi K3 (2.8 TRILLION params) on under 4GB→ works with almost every open modelThe biggest model on it needs the LEAST VRAM. K3 is sparse MoE, so it streams only the experts a token actually routes to instead of a whole dense layer.Link: https://lnkd.in/exD-jrwE2.8 trillion parameters running in less VRAM than a 70B.

Recently, I faced an interview for a #Generative #AI Engineer role, and the interview was heavily focused on RAG, LLMs, and production-grade AI systems.I thought of sharing some of the important RAG questions that were discussed during the interview. Hopefully, these will be helpful for anyone preparing for GenAI / AI Engineer / LLM Engineer roles. 🚀🔥 #Important #RAG Interview Questions🔹 What is RAG and why do we need it?🔹 RAG vs Fine-tuning — when would you choose which?🔹 Explain the complete RAG architecture.🔹 How do you decide chunk size and chunk overlap?🔹 What are embeddings and how do they work?🔹 How does vector similarity search work?🔹 How do you choose Top-K?🔹 What is Hybrid Search?🔹 What is Hybrid RAG and how is it different from Hybrid Search?🔹 Why do we need reranking?🔹 What is Query Rewriting and why is it useful?🔹 How do you improve poor retrieval quality?🔹 How do you reduce hallucinations in RAG?🔹 How do you handle questions when relevant information is not available in the knowledge base?🔹 How do you evaluate Retriever performance?🔹 How do you evaluate the final LLM response?🔹 How do you identify whether an issue is with Retrieval or Generation?🔹 How would you handle a large context window?🔹 How do you reduce latency in a production RAG system?🔹 How would you scale RAG from 100 queries to 100K+ queries?🔹 How would you scale a vector database from 10K to 1M+ documents?🔹 Where would you use caching in RAG?🔹 How do you monitor a RAG system in production?🔹 What are common security risks in RAG?🔹 What is Prompt Injection and how do you prevent it?🔹 How do you implement Guardrails in a RAG application?🔹 How would you design a multi-tenant RAG system?💡 Key takeawayModern GenAI interviews are going beyond just asking “What is RAG?”Interviewers are increasingly focusing on how you would build, evaluate, optimize, secure, monitor, and scale RAG systems in production.If you're preparing for a GenAI / LLM / RAG Engineer interview, I hope this list helps you identify the areas you should focus on.I’ll be covering these questions with practical explanations and real-world architecture examples in upcoming posts.What RAG topic do you find the most challenging?#GenerativeAI #GenAI #RAG #LLM #ArtificialIntelligence #MachineLearning #AIEngineer #LLMEngineer #GenerativeAIEngineer #InterviewPreparation #TechInterview #Python #VectorDatabase #HybridSearch #LangChain #LangGraph

Microsoft just open-sourced one of the best AI data analysis tools out there.It’s called data-formulator. Connect any source, CSV, Postgres, Bigquery, even a live URL, then build charts with a mix of drag-and-drop and plain english. The AI agent writes the SQL and transforms underneath and hands you a chart you can actually edit, not a code dump.- drag fields onto x/y/color and it builds the chart- anchor a cleaned result so follow-ups don't drift back to raw data- branch any chart to explore a variation - connect live data with auto-refreshLink: https://lnkd.in/es58xgrjbring your own model key and it runs locally.

Attended a demo by Logan Markewich who walked through LiteParse a genuinely clever piece of engineering worth breaking down.What makes it different: there's no LLM involved at all. Just clean markdown or JSON across PDF, DOCX, PPTX, XLSX and scanned images, at roughly 2ms per page.For anyone building agent pipelines, that's the real value. Reading a document no longer requires a model call, which strips latency, cost, and a whole category of hallucination risk out of the ingestion step. Fast pipelines stay fast, and the agent gets usable structure immediately.RAG ingestion is the obvious application. The more compelling one, in my view, is real-time agent tool calls where speed and determinism are exactly what the workflow demands.

There are three kinds of people using AI at work.Two of them are about to get squeezed.One of them is about to get paid.Every major technology wave changes the price of execution.Tractors didn't eliminate farm labor. They collapsed its value.Spreadsheets didn't eliminate accounting. They killed the bookkeeping premium.𝗔𝗜 𝗶𝘀 𝗱𝗼𝗶𝗻𝗴 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝘁𝗼 𝗰𝗼𝗴𝗻𝗶𝘁𝗶𝘃𝗲 𝘄𝗼𝗿𝗸.The model writes the memo. The model drafts the strategy. Execution becomes cheap.And when execution becomes cheap, value moves upstream. Toward the people who decide what problems matter, which systems get built, and where intelligence gets deployed.This isn't a technology shift. It's a leadership shift.Because the real divide isn't tools. 𝗜𝘁'𝘀 𝗹𝗲𝘃𝗲𝗿𝗮𝗴𝗲.1️⃣ Operator - uses AI to execute faster.Competes on 𝘀𝗽𝗲𝗲𝗱.2️⃣ Architect - builds systems that automate execution.Competes on 𝗱𝗲𝘀𝗶𝗴𝗻.3️⃣ Orchestrator - decides what work happens at all.Competes on 𝗷𝘂𝗱𝗴𝗺𝗲𝗻𝘁.Same tools. Completely different outcome.Where does value move when intelligence becomes cheap?The leaders who understand this will redesign how work happens.Everyone else will keep competing on execution. The cheapest layer in the stack.Which layer are you actually operating in most days?Operator, Architect, or Orchestrator?From Pilots to Platforms_________________________Want to build an Agentic Organization?Join our AI-native leadership cohort.It's a 4-week program for business leaders. Live sessions built around real challenges from your own organization, not toy demos. You leave with working agents, playbooks, and a transformation roadmap you can actually use to drive real outcomes.

London, New York, LA and San Francisco offer the WORST value on earth 💸Valerii Emelianov mapped 100 cities on cost of living vs quality of life using Numbeo data:▫️ Migration magnets rank worst: high costs, expensive housing, services that stop matching expectations▫️ Porto, Valencia and Prague have held the top of the value ranking for 7 years, and all three keep drifting right as costs climb▫️ Europe has the widest spread on the planet, from Switzerland to the Balkans▫️ Sun Belt and Pacific Coast cities are climbing fast on domestic migration▫️ Some cheap winners are pockets of comfort for elites and tourists rather than the median residentQuality of life here means purchasing power, safety, healthcare, housing, traffic, pollution and climate

Link copied to clipboard. View post.HomeMy NetworkJobs1Messaging16NotificationsMeFor BusinessTry Premium for ₹0Kaaviyasri Varshini RGenAI Builder @ KactiiChennai, Tamil NaduMeenakshi College Of EngineeringGrow your career with PremiumTry Premium for ₹0ConnectionsGrow your network12Feed postPriyanka Vergadia • 3rd+#1 Visual Storyteller in Tech | VP Level Product & GTM | TED Speaker | Enterprise AI Adoption at Scale | 250K+ CommunityBook an appointment3d • Follow👩‍🏫 I've been teaching in some form since I was 14. The biggest lesson it's taught me is: 𝗘𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝗹𝗲𝗮𝗿𝗻𝘀 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁𝗹𝘆.Which is actually great news right now. There's no shortage of AI courses out there. Different formats, different depths, different styles. You get to pick what fits how you learn.🤔 𝗕𝘂𝘁 that raises a 𝗕𝗜𝗚 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻: with so many options out there, including ones built with AI in a weekend, why choose my course?Three reasons, and none of them are the curriculum. 1️⃣ 𝗧𝗵𝗲 𝘀𝘁𝗼𝗿𝗶𝗲𝘀. What AI can't generate is real lived stories! A decade of bringing AI products at Google, Microsoft to millions of developers, or leading AI transformation for some of the biggest companies in the world. You're not paying for slides. You're paying for the war stories behind them.2️⃣ 𝗧𝗵𝗲 𝗰𝗿𝗲𝗱𝗶𝗯𝗶𝗹𝗶𝘁𝘆. Experience is scar tissue. I've been in the boardroom and in the trenches leading AI strategy to change management for companies like Chevron, UBS, Zurich Insurance, Barclays, etc getting thousands of developers to actually adopt AI.3️⃣ 𝗧𝗵𝗲 𝗰𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝘆. My courses pull in everyone from senior leaders to individual contributors, all in the same room, actually talking to each other. Connecting right hires to right hiring managers. And that third one matters the most. In the AI age, 𝘁𝗵𝗲 𝗼𝗻𝗲 𝘁𝗵𝗶𝗻𝗴 𝘁𝗵𝗮𝘁 𝘀𝘁𝗮𝘆𝘀 𝗲𝗻𝘁𝗶𝗿𝗲𝗹𝘆 𝘆𝗼𝘂𝗿𝘀 𝗶𝘀 𝘁𝗵𝗲 𝗻𝗲𝘁𝘄𝗼𝗿𝗸 𝘆𝗼𝘂 𝗯𝘂𝗶𝗹𝗱. Not the certificate! That's what I'm building, one cohort at a time.👍 Want to be part of the community? I would love to have you for the next cohort starting in a few days with 30% discount EARLYBIRD: https://lnkd.in/gZAw3Y94👇 Photo: See what some of the students had to say about the course.#AIEducation #LinkedInLearning #TechnicalStorytelling #AITransformation #CommunityBuilding #CareerGrowth #LeadershipLessons

Stop optimizing tokens. Index your context instead!Retrieval quality is the foundation of context engineering. Karpathy described it best: the heavy cognitive work should happen at ingestion, not at query time. When knowledge is properly structured before retrieval, the model's job becomes reasoning, not sorting.Most AI systems focus on compressing what the model sees. The more important problem is what gets retrieved before the model sees anything.Token efficiency starts at retrieval. When context is properly indexed and prepared, the model spends its tokens on reasoning. When retrieval is weak, the model spends those same tokens sorting through noise and filling gaps from its own weights.That's the silent failure mode. Retrieval returns topically correct but incomplete context. The model completes the gaps from parametric knowledge and streams it out the same way as grounded content. No signal in the output tells you which parts came from retrieved context and which came from weights. The answer looks confident. It just isn't complete.The common assumption is that hallucination is the main failure mode. It's not. Models handle off-domain questions reasonably well now. If nothing in the retrieved context looks relevant, there's no material to build an answer on. The harder failure is partial coverage. The right document was retrieved. But not the full picture.Coverage gaps don't produce error messages. They produce confident answers with pieces missing.This gets worse when sources stay isolated. The same person might appear across multiple tools and systems. If those sources are indexed separately, the model has to figure out they refer to the same entity on its own. That's work that should happen before the model starts reasoning.Glean's system of context is built around this problem:• Unified index across all connected applications, not each source kept separate• Specialized indexes for different types of information: company data, code, experts, profiles, tools, and calendars• Multiple retrieval methods - semantic when meaning matters, lexical when exact terms matter, structured when fields and relationships need to stay intact• Enterprise Graph that maps relationships across people, teams, customers, and projects so relevance reflects how the company actually works• Memory that carries context forward across sessions and tasks• Tools that let the model act on what it findsThe gap between finding information and understanding it is where most AI systems fall short.I've shared the link in the replies!

The University School of Computing, Rayat Bahra University, Mohali successfully organised the IEEE International Conference on Secure Information Systems and Technologies (ICSIST 2026) from 17–19 August 2026 at Rayat Bahra University.The conference provided a vibrant international platform for researchers, academicians, industry professionals, scientists, and practitioners to exchange ideas and present cutting-edge research in cybersecurity, secure information systems, artificial intelligence, blockchain, IoT, cloud and edge computing, digital forensics, privacy, and emerging technologies.The event featured keynote and invited talks, technical paper presentations, expert sessions, knowledge exchange, and opportunities for international research collaboration, bringing together distinguished experts from academia and industry across the globe.I would like to thank our program chair, organising chair, keynote speakers, session chairs, technical programme committee, speakers, reviewers, authors, delegates, volunteers, and the entire organising team for their valuable contributions in making ICSIST 2026 a successful academic event.🌐 IEEE ICSIST 2026📅 17–19 August 2026📍 Rayat Bahra University, Mohali, Punjab, IndiaCongratulations to the entire University School of Computing, Rayat Bahra University team! 👏Dr. Sahil Verma, B.T, M.T, Ph.D.(C.S.E) Postdoc-3, SMIEEE,ACM,IAENGProf. Nithesh Naik#ICSIST2026 #IEEE #RayatBahraUniversity #UniversitySchoolOfComputing #CyberSecurity #SecureInformationSystems #ArtificialIntelligence #EmergingTechnologies #Research #Innovation #InternationalConference #AcademiaIndustryCollaboration #IEEEConference #RBU

Every single morning for the past 3 years, I would open my eyes & the next thing would be to check LinkedIn, & suddenly 3 months went by and I haven’t written a single post. The last 3 months, I was dealing with a lot all at once, and I’m still dealing with them. I had family emergencies because of which I was in India for 2 months, but thankfully, things are looking up now. Over the last few weeks, I’ve also been quietly processing the loss of a life changing job offer due to the geopolitical situation, and another dream offer that I had to turn down due to personal reasons. Amongst all this, and few other unfortunate events in my personal life, I started believing I didn’t belong on LinkedIn anymore. Because I had nothing to teach anyone, no wins to share, no great opportunities to talk about, and no way to inspire or motivate anyone. Most days, I was barely able to motivate myself. I only had losses and failures to deal with. I have often seen that people choose to only talk about their wins here, and even they do talk about failures, they do so looking at them in the rear view mirror. So, in a way, it’s still a win, right? I’ve never come across a person who has willingly chosen to talk about their losses and failures while they were living through them. I haven’t done so either. Until now. Last few weeks, a few of you reached out to check on me, since you hadn’t read a post from me in a long time, and I can’t tell you how much that simple gesture meant to me. Social media, its attention and the visibility it brings you is so fleeting, that if you disappear, you’re done. I risked all that by quietly disappearing, so I’m not sure if I still have the attention of all of my dear readers whose support I painstakingly garnered over 3 long years. But, here’s something: I’m still dealing with a lot, but I’m choosing to now share them while I’m at it, so I can tell you all about my experiences. If you’re dealing with something similar, maybe it’ll give you the perspective that you’re not alone in your difficult headspace. Because sometimes, that’s all we need. Starting this week, I’ll write more honestly than I’ve ever written before, talking about what I’ve lost in the past few months, how I’m dealing with them and why I’m still choosing to show up here, despite only losing and failing at everything. Hoping all of you will continue to support me and keep reading. Nothing but love and profound gratitude for you all. And yes, I’m back, and I hope you missed me!

RAG has a problem: documents have relationships. Most RAG pipelines turn a knowledge base into chunks, embed them, and retrieve the closest ones. That works well for simple questions. But ask something like "Which customer decision was influenced by the findings from last quarter?" and suddenly similarity search starts to struggle. The answer may depend on several entities and relationships spread across different documents. That's where GraphRAG gets interesting. Semantica combines vector search with knowledge graphs, entity extraction, relationship extraction, temporal context, and agent memory. Instead of treating every chunk as an isolated piece of information, it builds a connected representation of the knowledge. The really interesting part is multi-agent systems. Multiple agents can work against the same context graph. A researcher can discover something, write it into shared context, and an analyst can immediately use it without passing the entire result through another agent. And this isn't just conceptual. Their benchmarks on a 118K-node graph report reducing node search from 24 ms to 0.004 ms, along with significant improvements in semantic deduplication and candidate generation. I think the interesting shift is this: RAG is moving from "find me some relevant text" to "help me navigate a knowledge system." That's a much more powerful primitive for AI agents.

In many LLM apps, users ask the same question more than once. Sometimes the wording changes, but the meaning stays the same. For example: - “How many people live in Paris?” - “What is the population of Paris?” Without caching, both requests may trigger a new LLM call. Semantic caching helps avoid that. It can reduce response time and LLM costs by reusing answers for exact or semantically similar questions. The approach in this article uses two cache layers: - Redis for exact question matches - A vector database for questions with similar meaning - The LLM only when neither cache has a useful answer If the LLM creates a new response, the system saves it in both caches for future requests. This helps reduce repeated LLM calls. It can also improve response time and lower costs for common or similar questions. I have explained the end-to-end flow, including cache hits, cache misses, LLM fallback, and the data stored in Redis and the vector database.

I changed Hermes Agent’s USER.md while a session was still running. The file changed immediately. The next model request didn’t. It still carried the old user profile. Then I compressed the session. Same session ID, but the prompt rebuilt with the new profile. This matters because “state changed” does not mean “the model saw the change.” In a production agent, that gap can create stale decisions, confusing debugging, and false confidence that a correction is already active. That experiment exposed a boundary I think matters well beyond Hermes: durable state ≠ current model context. The same thing showed up elsewhere. A nested AGENTS.md rule arrived through a tool result without rebuilding the system prompt. A loaded skill entered conversation history rather than the prompt snapshot. And provider-cache state was separate again. Part 2 of Hermes Agent Architecture traces those boundaries through prompt assembly, context files, skills, caching, and compression. I also published the reproducible lab and validation receipts for anyone who wants to test it on another provider, or Hermes release.

𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲𝗱 𝗮 𝗰𝗼𝗱𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁 𝘄𝗵𝗲𝗿𝗲 𝗲𝘃𝗲𝗻 𝘁𝗵𝗲 𝗮𝗴𝗲𝗻𝘁 𝗹𝗼𝗼𝗽 𝗶𝘀 𝗮 𝗽𝗹𝘂𝗴𝗶𝗻.MIT. 130k stars in three days.━━━━━━━━━━━━━━━DeepSeek Harness (dsh). One command to try it:npx @deepseek-ai/dsh webStarts a local web UI. Node.js is the only prerequisite.━━━━━━━━━━━━━━━One idea: everything is a plugin.→ Model adapters, tools, sessions, sandboxes, storage→ The UI itself→ The agent loopMost frameworks let you swap the model. The rest is hardcoded. Here you change any layer from config, without forking their source.━━━━━━━━━━━━━━━What actually stands out→ Full session replay — every prompt, tool call and result in an append-only log. Resume, fork, replay any run.→ Three run modes — full toolset, code-orchestrated tool calls, or minimal shell plus editor.→ Bring your own model. Self-hosted vLLM, a fine-tune, anything.━━━━━━━━━━━━━━━Two caveats→ v0.1 developer preview. DeepSeek says in capitals there will be breaking changes.→ Cordis, the runtime underneath, is not DeepSeek's. It is a cordiverse project that has powered Koishi for four years.Best detail: the repo root has a .claude folder and a CLAUDE.md. They built it using Claude Code.━━━━━━━━━━━━━━━The harness layer is now commodity. The differentiator moves to what you plug in — your tools, your context, your domain workflows.

Agent loops are great... in a demo. In production, you need a graph around them.Loops work in a demo. In production it fails in ways the loop can't see. The draft reads clean and the facts are wrong. The copy is fine and it answers a question nobody asked.The fix that actually holds up is treating production as a graph, with the loops inside it.Each specialist still iterates, but inside bounds. Research loops until it has a claim ledger with citations. An audience pass returns a reader contract, basically who this is for and what they will push back on. An example loop generates proofs and discards the ones that don't fit.Drafting only starts once all of that gets compressed into shared state. One agent writes, a separate critic checks it against the rubric. The maker never grades its own homework.The most underrated box in the whole system is the verification gate. It doesn't just pass or fail a draft. It routes by failure type. Weak copy goes back to drafting. A fact gap goes back to research. A scope gap goes all the way up to the orchestrator, because the words were fine and the goal wasn't.And nothing ships without a human signing off on the high-risk claims.Ultimately the loops are the easy part. The routing, the shared state and the recovery paths are where the reliability comes from.

DoorDash cut hallucinations in their Dasher support chatbot by 90% and severe compliance issues by 99%. The fix was not a better model. Here's how their team built the quality layer around it: 1. Why the Old System Capped Out The old flow relied on pre-built resolution paths, so only a small subset of issues could be resolved. The knowledge base behind it had three problems: the right article was hard to find, the useful part inside took time to locate, and every article was English while many Dashers prefer another language. 2. Retrieval Starts by Condensing the Conversation An issue is spread across several messages, so the system first condenses the conversation into the core problem. That summary searches historical data for the top N similar cases already resolved with KB articles, and each match maps to an article that goes into the prompt template. Summary quality gates everything downstream. 3. Two-Tier Guardrail, Cheap Check First They tried a sophisticated guardrail model first. Response times and token usage made it prohibitively expensive. Shallow check → sliding-window semantic similarity between response and article segments. A close match is less likely to be a hallucination. LLM evaluator → runs only when the shallow check flags. Sees the response, the KB articles and the conversation history, then judges groundedness, coherence, compliance. Every test must pass before a Dasher sees the response. 4. What LLM Judge Measures They manually reviewed thousands of transcripts, then categorised quality into five aspects: retrieval correctness, response accuracy, grammar and language, coherence to context, relevance to the request. LLM Judge is prompted with open-ended questions. High-frequency answers become the rules used for monitoring. A human team still reviews random samples. Key Takeaways 1. The guardrail is an architecture decision, not a prompt. Cheap check first, expensive check only on suspicion. 2. Latency is the cost. Generate, guard, sometimes retry. They route to human agents rather than wear it. 3. They found the failure categories by reading transcripts, then automated what they found. Reliability here came from everything wrapped around the model, not the model itself.

The Redis creator got tired of Python inference stacks. So he's writing one in C and yesterday he made a 37 GB model run in 2 GB of GPU memory. Byteidentical.antirez has spent the last two weeks building h3-metal: native inference for MiniMax-H3, a 33B video+audio diffusion model, on Apple Silicon. Pure C plus Metal shaders. No Python anywhere in the inference path. MIT licensed.The commit that landed yesterday is the one worth studying. SSD streaming for the DiT:➡️ Tracked transformer storage drops from 36.5 GiB to 2.0 GiB. The output is byte-identical to the full-residency path. ➡️ Two block slots alternate: a background reader pulls the next transformer block off SSD while the GPU executes the current one. Measured 13–14.6 GB/s from the internal drive.➡️ Darwin uncached reads, on purpose so the filesystem cache never holds a second copy of weights that will be read again anyway.On top of that, his int8 TensorOps path cut a full denoise from 36.3 to 19.3 seconds on M5 Max.Kimi K3 streaming off NVMe. Gemma 4 26B in a 2 GB budget. Now a 33B video model. Same conclusion three times in one month, from three independent codebases: weights don't need to fit in RAM. They need to arrive on time.SSD bandwidth is quietly becoming the new VRAM.

 New paper: Attention is All You Have.56,804 agent skills exist. Your model only has ~100 reliable trigger slots.Every install dumps a skill’s description into the permanent system prompt. They all fight for the same scarce attention. The long tail dies. Your team’s private playbooks compete with random public ones. Token tax every message. Dilution. Reliability collapse. Installation was never the right primitive. It bundled three things that don’t belong together: content, persistence, and auto-triggering. Only the last one needs to live in the prompt.The new skill protocol separates them.- A path addresses any skill, subtree, or collection.- Reading it is using it. Zero residency.- :save vendors a clean copy into your project’s git tree so you own and adapt it.- :install adds one .gitignore-style line. That’s the only thing that costs prompt space.Any agent that can read files + run commands becomes a client with a single instruction file. Optional free hub for search + ranking across the entire corpus. Install less. Use more.Attention is all you have.A collaboration between AdaL and The University of Texas at Austin (Atlas Wang)

The biggest problem with Claude Code has just been solved.Everything your agent learns from your code gets wiped out when you close the session.And tomorrow it learns it all over again. And your teammate learns it from scratch again with their own agent. And again. And again.Nobody on your team is saving that work anywhere.This turns it into files:→ Generates a repo map in markdown→ Saves to your project→ Updates only when the code changes→ You share it via git and it benefits the whole team→ Runs locally, no accounts or keys neededA new hire reads the docs once. Your agent hasn't read anything in months.It's called Graft. Open source, MIT license.Link: https://lnkd.in/eNhr5f7z