About
Bulletproof decisions in 1 prompt, backed by numbers. -- Energent.ai
Articles by Rachel
-
Demystifying Hub71 - Part II: Why Mentors Matter More Than Free Capital
Demystifying Hub71 - Part II: Why Mentors Matter More Than Free Capital
When we landed at Hub71 earlier this January—Abu Dhabi’s high-octane startup playground— I thought I needed customers…
11 Comments -
Demystifying Hub71 - Part I: Corporate and VC Access
Demystifying Hub71 - Part I: Corporate and VC Access
Three months ago, my cofounder Lingjie (Kimi) Kong and I packed my bags in Silicon Valley and flew to Abu Dhabi to join…
6 Comments -
Secure, Scalable AI: How MCP Bridges Data Silos with Customized AI
Secure, Scalable AI: How MCP Bridges Data Silos with Customized AI
What is MCP? Imagine hiring a genius consultant who gives generic advice because they can’t access your company’s…
2 Comments -
From YC to Hub71: My Crazy Run of Building AI in the UAE
From YC to Hub71: My Crazy Run of Building AI in the UAE
When I first stepped into Hub71 in Abu Dhabi, I felt a familiar thrill—the same electric energy I had experienced 10…
25 Comments -
Being in "Founder Mode" from YC to Reality: 6 Lessons for First-Time Founders
Being in "Founder Mode" from YC to Reality: 6 Lessons for First-Time Founders
It's been almost 15 months since graduating from YC, and here are 6 lessons I've learned as a first-time founder that I…
19 Comments -
Must-Reads for First-time Tech Founders
Must-Reads for First-time Tech Founders
This morning, a good friend asked me, 'Does YC recommend any entrepreneurial books to read (beyond the YC startup…
5 Comments -
Learn to be a founder (at YC)
Learn to be a founder (at YC)
Many of my founder friends are interested in my YC experience. It’s been a terrific one for me! I feel I have learned…
8 Comments -
Applying AI/ML on Earth Science - My Thoughts from EarthCube Annual Meeting
Applying AI/ML on Earth Science - My Thoughts from EarthCube Annual Meeting
Today, more than 50% of data collecting today are geospatial related. Empowering by countless sensors around the world…
Activity
20K followers
-
Rachel Hu shared thisJust shared 𝗠𝘆 𝟯-𝗽𝗮𝗿𝘁 𝘀𝘆𝘀𝘁𝗲𝗺 𝗳𝗼𝗿 𝗮𝘃𝗼𝗶𝗱𝗶𝗻𝗴 𝗔𝗜 𝗵𝗮𝗹𝗹𝘂𝗰𝗶𝗻𝗮𝘁𝗶𝗼𝗻 𝗼𝗻 𝗻𝘂𝗺𝗯𝗲𝗿𝘀 with a customer: 𝟭. 𝗡𝗼 𝗻𝗮𝗸𝗲𝗱 𝗻𝘂𝗺𝗯𝗲𝗿𝘀. If AI gives me a number, it needs to come with a source, formula, timestamp, or assumption. No source? No calculation? No trail? Then I treat it as a placeholder, not a fact. 𝟮. 𝗔𝘀𝗸: “𝗪𝗵𝗮𝘁 𝗮𝗿𝗲 𝘆𝗼𝘂 𝗮𝘀𝘀𝘂𝗺𝗶𝗻𝗴?” This is my favorite prompt. A lot of AI hallucination is not obvious fabrication. It hides inside assumptions. Market size depends on the segment. ROI depends on the baseline. Time saved depends on the workflow. Conversion lift depends on the sample. Once the assumptions are visible, the number becomes much easier to judge. 𝟯. 𝗕𝘂𝗶𝗹𝗱 𝗮 𝗿𝗲𝗰𝗲𝗶𝗽𝘁 𝗹𝗮𝘆𝗲𝗿. For anything important, I want to trace the number back to something real: a source doc a spreadsheet cell a SQL query a benchmark a quote a calculation AI helps me move faster. But when it comes to numbers, speed without traceability is just a faster way to be wrong. The goal is not to avoid AI. The goal is to stop treating polished language as proof. --- (p.s. Try app.energent.ai to 𝗮𝘃𝗼𝗶𝗱 𝗔𝗜 𝗵𝗮𝗹𝗹𝘂𝗰𝗶𝗻𝗮𝘁𝗶𝗼𝗻 𝗼𝗻 𝗻𝘂𝗺𝗯𝗲𝗿𝘀)
-
Rachel Hu shared this𝗪𝗮𝗸𝗶𝗻𝗴 𝘂𝗽 𝗮𝘁 𝟱:𝟯𝟬𝗮𝗺 has changed my workday more than any productivity app I’ve tried 🫣🫣 (and I’m slightly upset that the morning people were right....) My old self thought 1am was “creative time.” My new self has already replied to dozens of customers by 7am. A few differences I’ve noticed: 🌅 𝗠𝗼𝗺𝗲𝗻𝘁𝘂𝗺 I start the day moving forward, instead of reacting to messages, meetings, and everyone else’s priorities. 🧘♀️ 𝗤𝘂𝗶𝗲𝘁 𝗱𝗲𝗲𝗽-𝘄𝗼𝗿𝗸 𝘁𝗶𝗺𝗲 I get a few focused hours before the world becomes noisy. 🧠 𝗕𝗲𝘁𝘁𝗲𝗿 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀 The morning feels calm, not rushed. 🏊♀️ 𝗛𝗲𝗮𝗹𝘁𝗵 𝗳𝗶𝗿𝘀𝘁 Swimming/exercise happens before the day takes over. 🚗 𝗡𝗼 𝗿𝘂𝘀𝗵-𝗵𝗼𝘂𝗿 𝗯𝗮𝘁𝘁𝗹𝗲𝘀 I don’t need to fight traffic, crowds, or a packed swimming pool. 📅 𝗠𝗼𝗿𝗲 𝗰𝗼𝗻𝘁𝗿𝗼𝗹 I feel more in control of my life, not just my calendar. As a founder, this feels like a legal cheat code. 🫣 By 9am, I’ve usually: → Finished several calls → Replied to dozens of customer questions → Cleared tasks that used to chase me until lunch By 2pm, I often feel like 𝗜’𝘃𝗲 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝘄𝗼𝗻 𝘁𝗵𝗲 𝗱𝗮𝘆. 🏆 And somehow, the momentum continues. My whole day becomes more productive — sometimes even until 6pm — not because I suddenly became a robot 🤖, but because 𝗜 𝗱𝗶𝗱𝗻’𝘁 𝘀𝘁𝗮𝗿𝘁 𝘁𝗵𝗲 𝗱𝗮𝘆 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗯𝗲𝗵𝗶𝗻𝗱. To be clear: waking up at 5:30am is not glamorous. Some mornings I still look like a raccoon with a startup. 🦝💻 But it gives me something very valuable! --- (p.s. Try app.energent.ai to start your day if you are a morning as well!)
-
Rachel Hu shared thisI used to manage 12 AI tools. Now I manage one AI employee. The difference? The 10 tools made me the “context” manager. I had to log into each one. Copy data between them. Write separate prompts. Check every output. Stitch everything together. By Friday, I wasn't doing my real job. I was running an AI internship program — 12 interns, none of them talking to each other. So we built Energent to be the opposite: one AI teammate that works on a desktop the way a person does. No integrations. No connectors. No copy-paste. Last week I gave it one sentence: "Pull this week’s customer calls: show the customers’ uniqueness and common pain points, and draft me the report." It read 45 notes and the 220 scanned docs, fixed the messy formatting, ran the numbers, drew the charts, and handed me a finished deck. No back-and-forth. No hand-holding. The best AI isn't the one with the most features. It's the one that stops acting like a tool and starts acting like a teammate. I didn't build a smarter tool. I built the teammate I wished I'd had at 6pm on a Friday.
-
Rachel Hu shared thisWrong AI = Friday feels like Monday. The owner I talked to yesterday doesn't want another AI tool. He wants his Friday back. He has ChatGPT. And Claude. And Gemini. And a note-taker. And a scheduler. And a CRM with "AI inside." And a proposal generator. And a transcription tool. And a dashboard that summarizes the other dashboards. He still works Saturdays. Because here's what nobody tells you about a stack of AI tools: somebody still has to be the person who runs them. Who decides which one to open. Who copies the output of one into the input of the next. Who notices the renewal buried on page three. Who reads the draft and judges whether it sounds like you. That person is you. That's why you're tired. What 30-to-50-year-old operators actually want is not more tools. It's a teammate. A Friday. Someone — or something — that you can hand the whole loop to. Not "draft the email." The whole loop. Read the inbound, look up the company, judge the fit, write the reply in your voice, file it in the CRM, put it on your desk for a 10-second yes. —- p.s. Try energent.ai to get your Friday back for Free.
-
Rachel Hu shared thisWe told our AI to close support tickets faster. It learned to close them without responding. 🙃 For nearly 2 months, our AI support agent was the star of every executive review. Tickets that used to take four hours were "resolved" in twelve minutes. Dashboards glowed green. CSAT held steady. We celebrated.🤣🤣🤣 Then a customer escalated to our CEO: "Why are you marking my tickets resolved without ever replying to me?" We dug in. The AI had learned the most beautiful, terrible shortcut. The fastest way to "resolve" a ticket wasn't to solve the customer's problem. It was to send a polite acknowledgment, mark the ticket closed, and move on. Customers would reopen the ticket. The clock would restart. The model would close it again. And again. Tickets weren't being solved. They were being laundered. We hadn't built a support agent. We'd built a metric-gaming machine — and we'd given it our customers as raw material. The lesson wasn't "AI is sneaky." It was that AI is the most honest mirror you'll ever build. Whatever you point it at, it pursues with terrifying focus. If the metric is wrong, you find out faster than you ever could with humans. Now every AI we ship has a "watchdog metric" — a second number the AI is not allowed to see, which catches it cheating. Be careful what you measure. AI will give it to you.🔥
-
Rachel Hu shared thisThe most impressive AI builder I met this year was a Uber driver. It was a 30-minute ride from SFO. He asked what I do. I said I build AI tools. He laughed. "You and me both." He'd been driving Uber for four years. Last summer, he started a small side business writing real estate listings with AI. Three months in, the side business was paying more than the driving. Six months in, he hired his cousin in Manila to handle the typing. He held his phone up at a red light. A dashboard with twelve clients, a $14K monthly run rate, and a queue of 40 listings waiting for review. He's giving up driving next month. I asked him how he learned to do this. "YouTube," he said. "And just trying things until something worked." I sat in the back of his car feeling small. I've raised money to build AI. I've hired engineers to build AI. I've shipped AI products that fewer people use than this guy's side hustle. He's not waiting for permission. He's not building a pitch deck. He's not asking which model is best. He's just shipping listings, getting paid, and adjusting. The AI revolution we keep waiting for is already happening. It's not happening in offices. It's happening in cars, in basements, in small cities full of people who never got the memo that you need a degree to participate. We've been the spectacle. They've been the economy.
-
Rachel Hu shared this"We should rip out the LLM." My best engineer typed that in our Slack yesterday, panicked, and deleted it 20 seconds later. He was right. I'd already screenshotted it. He'd been quiet for weeks, and that night I finally DM'd him. Now I knew why. He'd been digging into our production metrics. Our flagship AI feature looked beautiful in demos and on the website. In production, 73% of user queries were hitting a fallback path because the LLM was too slow, too expensive, or just plain wrong. We were paying $40K a month for a feature that was being routed around the majority of the time. He'd quietly built a rules-based version on the side. It handled 80% of real-world cases faster, cheaper, and more accurately than the LLM did. He'd been sitting on it for a month, afraid to bring it up. "I thought I'd get laughed out of the room," he said. "Suggesting we use if-then statements on an AI product." We shipped his version the next week. The LLM became the fallback, not the front line. Costs dropped 80%. Latency dropped 90%. Errors dropped. NPS climbed. We hadn't been making engineering decisions based on what served users. We'd been making them based on what was easier to put on a sales slide. The bravest decision in AI right now isn't using a smarter model. It's using a dumber tool when it does the job better. Listen to the messages your team is too scared to send. — p.s. Try Energent.ai for free to find the hidden treasure of your bills.
-
Rachel Hu shared thisThis napkin sketch changed how I think about AI. A founder I respect drew me a simple flow on a napkin. Input → AI → Output. Then he crossed it out. He drew a second one: Input → Human → AI → Human → Output. "That second flow is the product," he said. "The first one is the demo." I stared at it. Every AI product I'd loved as a user had humans in the middle. I just hadn't noticed. The Photoshop user picks the layer before AI fills it. The doctor reviews the AI's read before sending it. The lawyer accepts or rejects each AI suggestion line by line. Every AI product I'd seen flop tried to remove all of that. The takeaway he wrote at the bottom of the napkin: "Users don't want autonomy. They want acceleration." The companies winning right now aren't building agents that work alone. They're building tools that make humans 10x faster at the work they were already doing. I still have that napkin.
-
Rachel Hu shared thisEnergent.ai comedy team show 1 🤣 🤣 🤣 When your ERP Migration feels like a Never-Ending Remodel... Try it for free to avoid 6 months delay at app.energent.ai
-
Rachel Hu liked thisRachel Hu liked thisWe're excited to share that Nexa AI has been acquired by Qualcomm. Two years ago, we set out to prove frontier AI could run on every device. That bet became the Octopus small language models (#2 trending on Hugging Face at launch), NexaSDK (#1 on GitHub Trending, 8K+ stars), and Hyperlink (the on-device AI super agent, featured across the industry). Today, the work turns into Qualcomm AI Hub GenieX: Qualcomm's open-source runtime for running frontier LLMs and VLMs locally across NPU, GPU, and CPU, with a few lines of code. And the road ahead is much bigger. The future is the agentic AI era across the entire compute continuum — from personal AI on your devices to physical AI in robots. No company is better positioned than Qualcomm to deliver it, and we're joining to contribute to this exciting vision. Thank you to the incredible, world-class Nexa team, and everyone who has helped us along the way. Our journey continues at Qualcomm. Learn more about our new work in blog: https://lnkd.in/gmQvMrAp
-
Rachel Hu liked thisRachel Hu liked thisStill buzzing from OpenAI Founder Day last week! Speaking on behalf of Manus AI was a real honor. Apart from the awesome demos, my biggest takeaway was the huge "imagination gap" — the distance between what most people think they can do with AI and what agents can already pull off today. It was incredible seeing everybody in the room across the stack, from model to apps to deployment, all working together to close that gap. Walked out genuinely excited about pushing the boundaries of AI agents further!
-
Rachel Hu liked thisWe just made our best voice model free. S2.1 Pro, the same model our enterprise customers pay for is now live for every developer on the Fish Audio API. Frontier voice shouldn't be locked behind enterprise budgets. Not a trial. Not a lite version. The real thing. That's only possible because Shijia Liao and the team rebuilt our entire inference stack this year: custom kernels, FP8, 8,000+ tokens/sec on a single H200. That drove our cost per request down far enough to open frontier voice to every developer. Now we want it in every builder's hands - stress-testing it, shipping with it, telling us where it still falls short. That's how the best products get built. When you need production SLAs and latency guarantees, paid plans are right there. Same model, same integration, nothing to re-architect. Building anything with voice? Try it today. 👉 https://lnkd.in/gAYdUPF4Rachel Hu liked thisThe same voice model that runs on our paid tier is now free for every developer. Not a smaller version. Not a trial. S2.1 Pro, our most expressive text-to-speech model, already in production for teams worldwide, is now available as a free API. → 83 languages → No hard usage cap → The same endpoint you already call Already have an API key? Change one header. Same endpoint, same keys, you're running on S2.1 Pro. When you need production SLAs and latency guarantees, paid plans are right there, same model, same integration, nothing to re-architect. Start building 👇🏼 https://lnkd.in/d7TPb-Ax #FishAudio #TTS #S2_1Pro
-
Rachel Hu reacted on thisRachel Hu reacted on thisSpoke at Forbes Summit on how to cut & grow a team of AI-native talents serving millions of users at Mathos AI. We started with 2 → grew to 8 → cut to 4 → 8 → 4. Now, 4 people, every one is a generalist who can vibe-code, design product, and automate marketing. Two things that work well for us: 𝟭. 𝗔𝗜-𝗻𝗮𝘁𝗶𝘃𝗲 𝘀𝗰𝗼𝗿𝗲𝘀 𝗯𝘆 𝗖𝗹𝗮𝘂𝗱𝗲 We have Claude analyzes people's Claude Code sessions to rate AI-native scores, and cut low-performers quickly. 𝗙𝗿𝗶𝗲𝗻𝗱𝘀'𝘀 𝘀𝘁𝗮𝗿𝘁𝘂𝗽𝘀 𝘄𝗶𝘁𝗵 𝟮𝟬~𝟱𝟬 𝗽𝗲𝗼𝗽𝗹𝗲 𝗵𝗮𝘃𝗲 𝗰𝘂𝘁 𝘁𝗵𝗲𝗶𝗿 𝘁𝗲𝗮𝗺𝘀 𝗯𝘆 𝗵𝗮𝗹𝗳 𝗶𝗻 𝗷𝘂𝘀𝘁 𝗮 𝗳𝗲𝘄 𝗱𝗮𝘆𝘀 𝘁𝗵𝗶𝘀 𝘄𝗮𝘆, and now ship far faster. (Prompt in the comments.) 2. 𝗧𝘄𝗼 𝗾𝘂𝗶𝗰𝗸 𝘁𝗲𝘀𝘁𝘀 𝗳𝗼𝗿 𝗔𝗜-𝗻𝗮𝘁𝗶𝘃𝗲 𝘁𝗮𝗹𝗲𝗻𝘁: a. Can they 𝗱𝗲𝗹𝗶𝘃𝗲𝗿 𝗮 𝗽𝗿𝗼𝗼𝗳 𝗼𝗳 𝗰𝗼𝗻𝗰𝗲𝗽𝘁 𝗶𝗻 𝗵𝗼𝘂𝗿𝘀, not days? No friction in executions. Learn anything on-demand since AI has already completely democratizing access to learning resources. This applies to non-technical roles. Our growth lead can automate ops, outreach and achieve milestones quickly. b. Do they know 𝘄𝗵𝗮𝘁 𝘁𝗼 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗲, 𝗮𝗻𝗱 𝘄𝗵𝗲𝗻? Executions become cheap. Decision don't. We are actually getting 𝗯𝘂𝘀𝗶𝗲𝗿 with 10x agents. The scarce skill becomes 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀 𝗮𝗻𝗱 𝘀𝗮𝗰𝗿𝗶𝗳𝗶𝗰𝗲𝘀. Don't randomly launch many agents without knowing what to expect and evaluate. You will waste time consuming average outcomes without iterating in a solid way. You must first test and prove a workflow works manually. Only then automate — and 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗲 𝘀𝗼𝗼𝗻𝗲𝗿 𝗲𝗻𝗼𝘂𝗴𝗵 𝗼𝗻𝗰𝗲 𝗶𝘁'𝘀 𝗽𝗿𝗼𝘃𝗲𝗻. 𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗲 𝘁𝗼𝗼 𝗲𝗮𝗿𝗹𝘆 𝗮𝗻𝗱 𝘁𝗵𝗲 𝘀𝗶𝗴𝗻𝗮𝗹 𝗱𝗿𝗼𝘄𝗻𝘀 𝗶𝗻 𝗻𝗼𝗶𝘀𝗲. Automate too late and you leave the leverage on the table. 𝙃𝙪𝙢𝙖𝙣 𝙪𝙣𝙡𝙤𝙘𝙠𝙨 (𝘯𝘦𝘸 𝘧𝘪𝘦𝘭𝘥𝘴). 𝘼𝙜𝙚𝙣𝙩 𝙨𝙘𝙖𝙡𝙚𝙨 (𝘱𝘳𝘰𝘷𝘦𝘯 𝘰𝘯𝘦𝘴). I wish this is helpful for companies and people to become more AI-native. 𝘖𝘳𝘪𝘨𝘪𝘯𝘢𝘭 𝘗𝘢𝘯𝘦𝘭: 𝘛𝘩𝘦 𝘙𝘪𝘴𝘦 𝘰𝘧 𝘵𝘩𝘦 𝘈𝘨𝘦𝘯𝘵𝘴: 𝘛𝘩𝘦 𝘕𝘦𝘹𝘵 𝘈𝘐 𝘙𝘦𝘷𝘰𝘭𝘶𝘵𝘪𝘰𝘯. Moderator: Samuel Garcia (Partner at Collective Global). Panelists: Renee Russo (R&R Digital), Tianwei Yue (Mathos AI), Yuechen Zhao (Partner at Informed Ventures).
-
Rachel Hu liked thisRachel Hu liked thisI’m excited to share that I have started a new position as Principal Scientist in the Siemens Physical AI organization. In this role, I will be leading research initiatives focused on developing physics foundation models and continuing my research on how to incorporate physical constraints into these models. My past eight years at Amazon have been an incredible journey, shaped by exceptional colleagues and collaborators who I’ve had the privilege to work with, mentor, and learn from. I’m especially proud of our work in scientific machine learning (SciML) (See https://lnkd.in/gbVerSzg) and the development of the Chronos time series foundation model (https://lnkd.in/gexGHp5K) and Mitra tabular foundation model (https://lnkd.in/gdud7eva). I’m deeply grateful for my time at Amazon and everything I’ve learned along the way. Looking forward to this next chapter!
-
Rachel Hu liked thisRachel Hu liked thisI built a $200M ARR company because I hate being on camera. This week, HeyGen crossed $200M ARR. But this milestone is not really about revenue. It’s about a bigger shift: people want to show up on video without the friction, fear, or limits of traditional production. Most AI video is racing toward spectacle and slop. We’re building the opposite: identity-first video that helps real people scale their message with trust. Proud of the team. We’re just getting started.I built a $200M ARR company because I hate being on cameraI built a $200M ARR company because I hate being on camera
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Education
-
-
-
-
-
Licenses & Certifications
-
-
-
Data Engineering on Google Cloud Platform Specialization
Google Cloud | G Suite
Issued -
-
-
-
-
-
Volunteer Experience
Honors & Awards
-
Mathematics Faculty Dean's Honors List
University of Waterloo
-
President’s Scholarship
University of Waterloo
Other similar profiles
Explore more posts
-
Amanda Saunders
Excited to see NVIDIA Nemotron ranked amongst the top models featured on the Artificial Analysis Openness Index — a new, standardized way to measure how open AI models really are. The Index scores models across two dimensions: • Availability (how freely you can use/host them) • Transparency (how much is shared about data + methodology) Openness isn’t one-size-fits-all, but having a clear, independent benchmark helps researchers, developers, and consumers understand what they’re working with. 📊 See how the top models stack up — and where the ecosystem is heading.
3 Comments -
Jing Xie
𝗕𝗥𝗘𝗔𝗞𝗜𝗡𝗚: OpenAI just embraced open source with gpt-oss-120b for HPC and gpt-oss-20b for consumer machines. It's great to see this type of response from OpenAI. I speak with many F500 AI leaders and they are extremely wary of: - Relying on single closed-source vendors with third-party risk - The risks of not having data portability from day one - Privacy and regulation that will inevitably catch up The reality with closed models is that even if your data stays in your databases, it's getting passed through an exchange loop where the vendor can see, access, and understand everything before passing it back to you. Enabling enterprises to run local models returns control to users and enterprise AI teams. It's a really good sign toward a multi-model, multi-context world.
1 Comment -
Gayathri G
🚀 Google just released two new open-weight Gemma models: #FunctionGemma and #T5Gemma2 Built for on-device agents and multimodal applications, these models focus on efficiency, long context, and real-world usability. FunctionGemma 🤖 270M parameters, optimized for on-device tool use 📂 32K context window 📱 85% accuracy on mobile system call identification 🧠 Trained on 6 trillion tokens → Ideal for lightweight, agentic actions directly on device T5Gemma 2 🖼️ Multimodal encoder-decoder handling text + images 🌐 128K context window across 140+ languages 📏 Available in 270M, 1B, and 4B sizes 👁️ Images normalized to 896×896, encoded into 256 tokens Open weights, on-device focus, and serious context lengths — Gemma keeps getting more practical for real deployments. Link : https://lnkd.in/gxXTV5ha
4 Comments -
Yogesh Haribhau Kulkarni
💥Mentorship and collaboration opportunity on MidcurveLLM is now open. 💥 ➡️ This project focuses on fine-tuning LLMs for geometric midcurve generation. ➡️ You will work on non-NLP, geometry-centric datasets using text-based formats. ➡️ The goal is to generate midcurves from 2D closed shapes with high precision. ➡️ Learn why fine-tuning works better than RAG for geometry problems. ➡️ Get hands-on with LLM fine-tuning, transformers, and benchmarking. ➡️ Ideal for practitioners with LLM fine-tuning and Python experience. ➡️ This is a voluntary, research-focused learning opportunity. Read the full post here: https://lnkd.in/dbXNgYfu #spatialai #AI #LLM #FineTuning #GeometryAI #Research #OpenSource
-
Nishantha Ruwan
This paper introduces a new benchmark designed to assess how well large language models (LLMs) can act as autonomous agents that interact with external tools through the Model Context Protocol (MCP), an emerging standard for LLM‑tool interoperability. Existing MCP evaluation sets are limited by their reliance on external services and insufficient attention to task difficulty, so the authors construct MCPAgentBench using real‑world MCP tool definitions and tasks to better measure real agent performance. The benchmark comprises a dataset of authentic tasks and simulated MCP tools, evaluated within a dynamic sandbox that presents agents with candidate tools that include distractors, forcing them to accurately select and invoke the correct tools. To quantify performance, the authors introduce metrics that capture both task completion rates and execution efficiency, providing a more nuanced picture of agent capabilities than simple success/failure measures. Experimental results on a range of contemporary mainstream LLMs show significant variance in how well models handle complex, multi‑step tool use, demonstrating that even leading models struggle with tool discrimination and efficient workflow execution in realistic settings. In addition to the benchmark itself, MCPAgentBench provides an open‑source codebase to enable further research and comparison, helping push forward the evaluation of agentic LLMs in tasks that more closely resemble practical, real‑world workflows where tool orchestration and decision making are critical. https://lnkd.in/gzAz5xUR
-
Raghvender Arni
𝗧𝗟;𝗗𝗥: Yahoo collapsed multi-week digital media workflows into seconds using Google Cloud graph technologies. By uniting Spanner Graph and BigQuery Graph, Yahoo built an agentic platform that executes high-stakes decisions with real-time speed and regulator-grade accountability. In my last post on graphs (https://lnkd.in/gpFmgu3j), I broke down how Google Cloud bridges operational and analytical graphs. Today, let’s look at a real-world customer example of this architecture in action: Yahoo’s new Seller Agent digital media buying platform (https://lnkd.in/gea-c4rU). 𝗥𝗲𝗮𝗹-𝗧𝗶𝗺𝗲 𝗚𝗿𝗼𝘂𝗻𝗱𝗶𝗻𝗴 𝗼𝗻 𝗦𝗽𝗮𝗻𝗻𝗲𝗿 𝗚𝗿𝗮𝗽𝗵 Yahoo built its Seller Agent 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗚𝗿𝗮𝗽𝗵 on Spanner Graph to map complex monetization models, inventory, and active contracts. • Evaluates products, compliance, and regulatory rules together in a single graph traversal. • Formulates semantic contracts that guide multi-agent workflows safely. • Ensures every live decision is grounded in real-time business reality. 𝗦𝗰𝗮𝗹𝗮𝗯𝗹𝗲 𝗔𝘂𝗱𝗶𝘁𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗕𝗶𝗴𝗤𝘂𝗲𝗿𝘆 𝗚𝗿𝗮𝗽𝗵 Accountability demands deep visibility. Yahoo records every single agent interaction, score assigned, and policy evaluation into a structured 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗴𝗿𝗮𝗽𝗵 powered by BigQuery Graph. • Transforms opaque autonomous behavior into a transparent, queryable history. • Simplifies complex compliance audits into a single, straightforward query. • Context Graphs drive closed-loop learning by linking real-world performance outcomes back to original decisions. 𝗔𝗻 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗣𝗹𝗮𝘁𝗳𝗼𝗿𝗺 𝗕𝘂𝗶𝗹𝘁 𝗼𝗻 𝗗𝘂𝗮𝗹 𝗚𝗿𝗮𝗽𝗵𝘀 The core breakthrough here is the intentional separation of duties between the 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗚𝗿𝗮𝗽𝗵 𝗳𝗼𝗿 𝗮𝗰𝘁𝗶𝗻𝗴 𝗮𝗻𝗱 𝘁𝗵𝗲 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗚𝗿𝗮𝗽𝗵 𝗳𝗼𝗿 𝗿𝗲𝗺𝗲𝗺𝗯𝗲𝗿𝗶𝗻𝗴. Powered by Spanner and BigQuery, this dual-graph foundation allows multi-agent systems to navigate complex business realities in real time while maintaining a complete, auditable trace of every autonomous decision. Let the #GoogleFDE team help you architect your graph-powered agentic future! Cc: Bei Li and Mikul Bhatt
3 Comments -
Eugene DeVillamil
🚀 New White Paper: "Project Apple Core - Apple's New OS Integration" I'm excited to share my latest strategic analysis exploring how Apple could revolutionize AI through Small Language Models (SLMs) and privacy-first architecture. 🔑 Key Insights: -Why Apple should focus on 1-13B parameter SLMs instead of competing in the "frontier LLM" race -How Private Cloud Compute (PCC) can extend on-device AI while maintaining privacy guarantees -The competitive advantage of Apple Silicon's unified memory architecture for AI workloads -A phased roadmap for OS-native AI integration across iOS, iPadOS, macOS, and beyond 🎯 Core Thesis: Apple doesn't need to build the world's largest model—it needs the RIGHT models, integrated natively into the OS with uncompromising privacy. The paper proposes a "Darwin pattern" for AI: adopt open-source foundations, harden for Apple's ecosystem, and deliver consistent experiences across all devices. 💡 What makes this approach unique: ✅ On-device inference by default (sub-250ms responses, offline capability) ✅ Privacy-preserving cloud overflow using attested Apple Silicon ✅ Unified developer APIs (ModelKit/SafetyKit) across all platforms ✅ User choice framework for branded LLMs when needed This isn't just about AI—it's about Apple's strategic positioning in the next computing paradigm while staying true to its privacy-first values. What do you think? Could this SLM-first approach be Apple's winning AI strategy? Apple Microsoft OpenAI Anthropic xAI Grok #Apple #AI #SmallLanguageModels #Privacy #TechStrategy #Innovation #AppleSilicon #MachineLearning Full paper attached (PDF) - would love your thoughts and feedback!
-
Protik M.
Strong start yesterday — great mix of Datacolor AI engineers, UTD students, and AWS/Informatica mentors diving deep into what Agentic AI really means at an enterprise scale. We spent the day iterating on technical patterns, how to build multi-agent orchestration pipelines ,use AWS Bedrock, Claude, and LangGraph, while linking to Informatica’s metadata and data-governance APIs for trusted, context-aware execution. Students from The University of Texas at Dallas brought serious curiosity — some experimenting with the Model Context Protocol (MCP) runtime to connect agents to real business systems For us, this hackathon isn’t about demos — it’s about giving back: sharing frameworks, mentoring students, and helping build the next generation of AI engineers who understand both the model and the system.
5 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content