Instagram story viewer> @viktoria.semaan> Posts

Viktoria Semaan

@viktoria.semaan

327

posts

308.4K

followers

146

following

👩🏻‍💻 Built AI systems at Databricks & AWS for Fortune 500s
🎬 AI Explained - how AI actually works
📩 Partnerships: [email protected]

POSTS STORIES REELS TAGGED

Download All

5 AI projects that will get you hired in 2026 ↓❌Stop wasting time vibe coding throw away simple apps. Learn enterprise skills that pay well.If you want to build practical AI projects that enhance your skills and create a portfolio that stands out, these 5 go-to projects .Not throw away demos. Real systems that handle: → Millions of data points → Enterprise security → Multi-modal inference → Persistent memoryIf you’re serious about landing a $160K+ AI job, build these:1️⃣ RAG from Scratch 2️⃣ Social Media Agent3️⃣ Medical Image Analysis4️⃣ MCP Tool-Calling Agents 5️⃣ Production AI MemoryEvery project has:✅ Open-source code✅ Production patterns✅ Real-world use cases🔗 Comment LINKS for all 5 project links.P.S Medical Imaging 3 shows real-world implementations and how you can handle scale + real world data.#AI #FreeTutorials #TechCareers #Databricks #TechCareer by @viktoria.semaan

3k

1k

5 months ago

Career Series 1 of 3: “AI Mastery”. So what level are you at right now?Most people get stuck between beginner and intermediate because they don’t have a clear roadmap. 🔗 Comment COURSE and I’ll send you a link to 4-part specialization that takes you from basics to building AI solutions.What is the biggest challenge holding you back? I’ll answer your questions in the comments.P.S. In the next video I’ll share resources I used to land offers for senior roles. ____Hello 👋 I’m Viktoria, Educator and Principal Technologist at Databricks. I share practical Al tips and resources. Follow for more educational content.#GenAIRoadmap #TechCareers #AIEducation #TechJobs #GenAI by @viktoria.semaan

3k

2k

4 months ago

Ever wonder why your phone doesn’t melt when you run a GenAI app? It’s because you aren’t just using 1 processor - you’re using an entire symphony of specialized silicon.In 2026, “compute” isn’t a monolith. Here is how the labor is actually divided:1️⃣ 𝗖𝗣𝗨: 𝗦𝗲𝗾𝘂𝗲𝗻𝘁𝗶𝗮𝗹 𝗣𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 Features powerful cores (typically 4-12) that handle diverse tasks with precision. It acts as the “General Manager,” managing system logic and complex branching that requires high single-thread speed.2️⃣ 𝗚𝗣𝗨: 𝗠𝗮𝘀𝘀𝗶𝘃𝗲𝗹𝘆 𝗣𝗮𝗿𝗮𝗹𝗹𝗲𝗹 𝗣𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 Contains thousands of smaller cores working simultaneously. This parallel throughput is what transformed AI training from a multi-week ordeal into a matter of hours.3️⃣ 𝗧𝗣𝗨: 𝗦𝘆𝘀𝘁𝗼𝗹𝗶𝗰 𝗔𝗿𝗿𝗮𝘆 𝗳𝗼𝗿 𝗧𝗲𝗻𝘀𝗼𝗿 𝗠𝗮𝘁𝗵 Google’s custom cloud chip. Data flows in waves, reusing weights hundreds of times to minimize memory access. It powers Search and Gemini at massive scale with extreme efficiency.4️⃣ 𝗡𝗣𝗨: 𝗢𝗻-𝗗𝗲𝘃𝗶𝗰𝗲 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗘𝗻𝗴𝗶𝗻𝗲The 2025 standard for edge computing. It runs AI locally—like Face ID and photo editing—preserving your privacy and battery life without needing the cloud.The future of AI isn’t just “faster” chips; it’s about architectural diversity.Share it with someone who is interested in learning AI.___👋 Hi, I’m Viktoria, Principal AI Evangelist at Databricks and Educator. I make complex tech accessible. Follow for more AI insights!#AIExplained #GenAI #LearnAI #Tech #TechCareers by @viktoria.semaan

5k

78

4 months ago

Telling AI “you are an expert” can make an LLM less accurate. A new study tested the most common prompting trick people use: assigning the model an expert persona. The result was surprising:✅ Expert personas helped with tone, writing, and safety. ❌ But on factual questions, every version they tested made accuracy worse.On a standard knowledge benchmark, it dropped from 71.6% to 68.0%.The trick everyone copies into their templates quietly backfires.This is part of why the conversation is shifting from prompt engineering to context engineering.Prompt engineering is about instructions. You tell the model how to behave: what role to play, what tone to take, what steps to follow. You write “act as a senior engineer” or “respond formally” and refine the wording until the output looks right.Context engineering is about information. Instead of telling the model how to act, you give it what it actually needs to work: your situation, your constraints, your data, examples of what good looks like, and what you want to avoid.The model already knows how to think. What it’s missing is your context.Replit Agent 4 is built around this. You describe the outcome, it plans, researches, and builds.What this looks like in practice:→ Put the outcome, constraints, and data in one brief, not piecemeal → Let the agent finish its plan before you step in → Edit elements directly on the Canvas, not in a new prompt → Skip reopening long chats for small fixes, it reprocesses everything🔗 Get free Replit credits with link in my bio.Follow for more practical AI tips.#promptengineering #contextengineering #replitpartner #aitips by @viktoria.semaan

346

15

9 days ago

What is the best advice you have ever been given? Share in the comments! by @viktoria.semaan

274

12

13 days ago

I’ve been struggling to produce infographics with AI that had no spelling mistakes. I kept taking the output and patching it manually in Canva. I finally figured out how.The problem is the model architecture.𝗗𝗶𝗳𝗳𝘂𝘀𝗶𝗼𝗻 𝗠𝗼𝗱𝗲𝗹𝘀Diffusion models work in cycles across the entire image at once. There is no concept of sequence or layout just repeated rounds of denoising everything globally.𝘗𝘳𝘰𝘮𝘱𝘵 → 𝘵𝘦𝘹𝘵 𝘦𝘮𝘣𝘦𝘥𝘥𝘪𝘯𝘨 → 𝘳𝘢𝘯𝘥𝘰𝘮 𝘯𝘰𝘪𝘴𝘦 𝘢𝘱𝘱𝘭𝘪𝘦𝘥 𝘵𝘰 𝘦𝘯𝘵𝘪𝘳𝘦 𝘪𝘮𝘢𝘨𝘦 → 𝘪𝘵𝘦𝘳𝘢𝘵𝘪𝘷𝘦 𝘥𝘦𝘯𝘰𝘪𝘴𝘦 𝘤𝘺𝘤𝘭𝘦 → 𝘥𝘦𝘤𝘰𝘥𝘦 𝘵𝘰 𝘧𝘪𝘯𝘢𝘭 𝘪𝘮𝘢𝘨𝘦→ Every cycle processes all pixels simultaneously, not in order→ The model cannot plan a heading, then place a table below it, then align columns→ Text and structure emerge from denoising noise, not from reasoning→ Every correction restarts the full cycle from scratch(NEW!) 𝗗𝗲𝗰𝗼𝗱𝗲𝗿-𝗢𝗻𝗹𝘆 𝗔𝘂𝘁𝗼𝗿𝗲𝗴𝗿𝗲𝘀𝘀𝗶𝘃𝗲 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗲𝗿Luma’s Uni-1 uses the same architecture as the large language models you already use. The image is first broken into tokens, and the model predicts each one based on everything that came before it.𝘗𝘳𝘰𝘮𝘱𝘵 → 𝘪𝘮𝘢𝘨𝘦 𝘵𝘰𝘬𝘦𝘯𝘪𝘻𝘦𝘥 𝘪𝘯𝘵𝘰 𝘥𝘪𝘴𝘤𝘳𝘦𝘵𝘦 𝘵𝘰𝘬𝘦𝘯𝘴 → 𝘱𝘳𝘦𝘥𝘪𝘤𝘵 𝘯𝘦𝘹𝘵 𝘵𝘰𝘬𝘦𝘯 𝘣𝘢𝘴𝘦𝘥 𝘰𝘯 𝘢𝘭𝘭 𝘱𝘳𝘦𝘷𝘪𝘰𝘶𝘴 𝘵𝘰𝘬𝘦𝘯𝘴 → 𝘳𝘦𝘱𝘦𝘢𝘵 𝘢𝘤𝘳𝘰𝘴𝘴 𝘧𝘶𝘭𝘭 𝘴𝘦𝘲𝘶𝘦𝘯𝘤𝘦 → 𝘥𝘦𝘤𝘰𝘥𝘦 𝘵𝘰𝘬𝘦𝘯𝘴 𝘣𝘢𝘤𝘬 𝘵𝘰 𝘪𝘮𝘢𝘨𝘦→ Text and image tokens are processed together in order, not in parallel cycles→ The model plans composition before it renders anything→ When it places a heading, it already knows where the columns go→ Edits are surgical only the specified region gets retokenizedI have been using Luma AI for my infographics and it has saved me so much time. I can edit specific elements without regenerating the whole image.Follow for more practical AI tips.#infographics #imagegeneration #aitips by @viktoria.semaan

217

16

15 days ago

For years, we said: “You need 𝘢 𝘭𝘰𝘵 𝘰𝘧 𝘥𝘢𝘵𝘢 to train great AI models.” That’s changing.I sat down with Rev Lebaredian, VP of Simulation at NVIDIA Omniverse, who runs Physical AI. His framing reshaped how I think about training:𝘍𝘰𝘳 𝘱𝘩𝘺𝘴𝘪𝘤𝘢𝘭 𝘈𝘐 𝘤𝘰𝘭𝘭𝘦𝘤𝘵𝘪𝘯𝘨 𝘳𝘦𝘢𝘭 𝘥𝘢𝘵𝘢 𝘪𝘴 𝘵𝘰𝘰 𝘦𝘹𝘱𝘦𝘯𝘴𝘪𝘷𝘦, 𝘵𝘰𝘰 𝘥𝘢𝘯𝘨𝘦𝘳𝘰𝘶𝘴, 𝘰𝘳 𝘴𝘪𝘮𝘱𝘭𝘺 𝘪𝘮𝘱𝘰𝘴𝘴𝘪𝘣𝘭𝘦. 𝘠𝘰𝘶 𝘤𝘢𝘯’𝘵 𝘵𝘳𝘢𝘪𝘯 𝘢 𝘴𝘦𝘭𝘧-𝘥𝘳𝘪𝘷𝘪𝘯𝘨 𝘤𝘢𝘳 𝘣𝘺 𝘦𝘹𝘱𝘰𝘴𝘪𝘯𝘨 𝘪𝘵 𝘵𝘰 𝘦𝘷𝘦𝘳𝘺 𝘦𝘥𝘨𝘦 𝘤𝘢𝘴𝘦 𝘪𝘯 𝘱𝘳𝘰𝘥𝘶𝘤𝘵𝘪𝘰𝘯.So instead of data, you use compute.The new pipeline: • Pre-train the model on real-world and internet data • Generate post-training data through simulation • Post-train the model on simulated scenarios • Evaluate with policy training • Feed it back into the simulatorThe flow: Data -> Pre-Trained Model ->Post-Training data -> Post-Trained Model -> EvaluationIn other words:more simulation cycles = more training iterations = better models.Curious, how do you see simulation and synthetic data changing the future of AI training? by @viktoria.semaan

413

5

17 days ago

This is for anyone who feels behind.Here are 5 things I learned in my 30s that I wish someone had told me sooner.1️⃣ Reinvent yourself as often as the world does.I started as an automation engineer. AI didn’t exist. Cloud didn’t exist. I couldn’t have imagined going from a factory floor in West VA to speaking on stages in Paris and New York. Stay curious. Never get too comfortable with who you are today.2️⃣ Plan long-term and invest early. Even if it’s small.When I came to the US, I wasn’t making much, but I invested every month. I skipped the designer bags in my 20s and built financial freedom that now lets me walk away from any role that doesn’t align with my values. That’s the real freedom.3️⃣ Prioritize family. Especially if you’re a woman.I joined tech in my mid-30s, after having 3 kids. For years I worried I was “behind.” I wasn’t. My kids are the accomplishment I’m proudest of. Careers can be rebuilt at 35, 45, 55. Some windows in life don’t reopen.4️⃣ Be kind. Especially to people “below” you.Early in my career, I watched leaders treat junior people like they were disposable. AI has reshuffled the deck. Some of those juniors are founders now. The hierarchy you see today won’t be the hierarchy in 5 years. Lift people up. When your network wins, you win with it.5️⃣ Stop comparing your real life to someone else’s highlight reel.Social media is a museum of everyone’s best 2 seconds. You’re comparing your Tuesday morning to someone else’s peak moment. Celebrate the hard things you’ve overcome. That’s where real growth lives.Grateful for every chapter that brought me here 🙏P.S. My birthday was 2 weeks ago. I guess I’m always a little late… by @viktoria.semaan

854

40

a month ago

388

36

a month ago

🔗 Comment VIKTORIA and I’ll send you a code to get 1 month for FREEPeople are building real businesses with zero coding background right now.Paulius Masalskas built CreatorHunter (an influencer database for startups) on his daily commute using AI tools. No team, $30K in revenue, 350+ paying users.Maor Shlomo bootstrapped Base44 to 250K users and $189K monthly profit in 6 months then sold it to Wix for $80M cash.His team was 8 people.This is what building looks like in 2026.@bubble__hq is where a lot of these ideas come to life 👇 Unlike AI code generators that hand you raw code to manage, Bubble gives you one visual environment: frontend, backend, database, and logic all in one place. Built for:◾ SaaS products & subscription apps ◾ Marketplaces ◾ Custom CRMs & client portals ◾ Internal dashboards & admin tools ◾ MVPs you can validate before writing a single line of codeGet a free month with code VIKTORIA1MO Build it. Build it. Test it. Ship it. 🚀#prototypetoprofit #vibecoding #aitools #sponsored #bubbleai by @viktoria.semaan

182

106

a month ago

Most people think AI just answers questions. Manus agents actually complete tasks for you. Launched March 2025, acquired by Meta in December 2025 - this is the autonomous agent that operates like a digital employee, delivering finished results while you are offline.What makes it different:🖥️ Manus Computer: real-time window showing every step the agent takes ⚡ Agent Skills: reusable workflows you build once and trigger on demand 🔄 Async operation: runs in the background, notifies you when done 🧠 Multi-model intelligence: combines Claude and Alibaba’s Qwen for each task 📅 Scheduled tasks: set it once, runs automatically every weekI used the stock analysis skill to build a weekly research workflow delivered to my inbox every Friday before markets open.🔗 Get 1,000 free credits with my link in bio. Have you tried building long running AI agents? Would love to hear about your experience! #aiagents #manus #learnAI by @viktoria.semaan

321

30

a month ago

Stop just asking AI. Start letting it execute for you. Let’s break down new capabilities ↓Accio Work turns AI into a full-time assistant: it plans, connects tools, and completes tasks so you can focus on what matters.🧠 Skills: Turn tasks into reusable instructions🤝 Connectors: Gmail, Notion, Slack, Instagram📱 Channels: Dispatch agents from your phone⏰ Scheduled Tasks: Pull & summarize industry news daily🤖 Model Choice: Gemini for images/videos, Claude for reasoning🔗 Comment ACCIO and I’ll send the link.#MyAccioWorks #Accio #AccioAgent #ClaudeCowork #LearnAI by @viktoria.semaan

372

76

2 months ago