# Karthick Arun > Founder of Setu Labs. Passed AWS's Cloud Practitioner exam at 9, now building Lucky, a voice-first AI guide for blind users, alongside AI research and student tools. Karthick Arun · Founder of Setu Labs. Gilbert, Arizona. Basha High School, Class of 2027. AWS turned me away from its Cloud Practitioner exam for being too young, so my dad and I emailed Andy Jassy. AWS made an exception, later dropped its age limit, and invited me to re:Invent. After teaching other kids cloud computing over Zoom, I founded Setu Labs to build Lucky, a digital guide dog that reads PDFs, email, and the web aloud. ## Right now - Founder & CEO, Setu Labs: Building Lucky, a voice AI for blind people, with Foundation Fighting Blindness chapters - AP Research: AI fraud detection: Fine-tuned a 20B open-weight model that beat GPT-5 on fraud detection across seven domains: 0.78 F1 vs. 0.65, and 16× faster - President, Basha DevOps Coding Club: Teaching younger students to code, something I started in 2020 ## Work What I’ve founded, researched, and shipped. Each project has its own page: why I started it, what I did, and how far it’s gotten. Highlights: - Lucky, a digital guide dog for the web: A family member lost their vision, and I saw how much their screen reader couldn't read. So at Setu Labs, I built Lucky: you talk, and it reads PDFs, email, and the web back to you in the right order. I've now presented it to 300 blind users at Foundation Fighting Blindness — their feedback shapes every release. - A small model beat GPT-5 at catching fraud: I fine-tuned two open-weight models on seven fraud domains. My hypothesis was wrong: the smallest model won, with 0.78 F1 vs. 0.65 for GPT-5, and it answered 16 times faster. - SnakeID tells you if a snake is venomous: Take a photo, and AWS Lambda plus OpenAI Vision return the species, a color-coded safety class, and one-tap emergency calls. District runner-up, with Special Congressional Recognition. ### Lucky A voice-first digital guide dog that reads PDFs and email aloud for blind and low-vision people. (https://karthick.me/work/lucky) - When: 2024 – now - Role: Founder & CEO, Setu Labs - Result: Exhibited at VISIONS 2026; presented to ~300 blind users - Stack: Next.js, LiveKit, Deepgram, OpenAI, AWS (Lambda, Bedrock, ECS) - Live: https://lucky.setulabs.ai Why I started: - A family member lost their vision. Watching them use a screen reader, I saw how much was locked away: screen readers read everything in order in the same flat voice, page numbers and figure labels included. Untagged or scanned PDFs often don't read at all. - At Setu Labs, I set out to build the tool I wanted them to have. What I did: - I lead product development and manage a team of graduate students. - With Lucky, you hold Space and talk. It reads PDFs in the right order with page-by-page navigation, handles Gmail by voice, and searches the web with cited sources. You can change its voice and speed just by asking. - The web app connects over LiveKit to a Python voice agent. PDFs go through an AWS backend of about 25 Lambdas, which recovers reading order and writes summaries. The agent never holds the user's login token: the browser makes every authenticated call. Progress so far: - I presented Lucky to about 300 blind users at a Foundation Fighting Blindness chapter meeting and recruited testers. Voice and speed came up, so now you can change both just by asking. - We co-develop Lucky with FFB chapters, and I exhibited it at VISIONS 2026. We'd priced Lucky at $20 a month, like most software, but visitors asked about cost: many blind users live on fixed incomes and rely on state agencies to buy their tools. What's next: - Take Lucky from booth demos to daily use, starting with the FFB chapters that already test it, and work out grants and state programs to reach people who can't pay $20 a month. ### Small models vs. GPT-5 on fraud detection Can a fine-tuned open-weight model beat a frontier model at spotting deception? Across seven domains, the smaller one did. (https://karthick.me/work/fraud-research) - When: 2025 – 2026 - Role: Independent researcher, AP Research - Result: A fine-tuned 20B model beat GPT-5: 0.78 vs. 0.65 F1 - Stack: GPT-OSS-20B, GPT-OSS-120B, GPT-5, Fireworks.ai, Python Why I started: - Everyone assumes the biggest model wins. For something like fraud filtering, which runs on every message, I wanted to know whether that's true, and what it costs. What I did: - I fine-tuned GPT-OSS-20B and GPT-OSS-120B for one epoch each on Fireworks.ai, then tested them against GPT-5 with no fine-tuning. - The data came from the DiFrauD benchmark: seven domains (fake news, SMS spam, phishing, Twitter rumours, fake product reviews, job scams, and false political statements) and 9,589 test examples. I scored precision, recall, F1, accuracy, and latency, then wrote it up as a full research paper. Progress so far: - The smallest model won: 0.78 mean F1, against 0.69 for the 120B model and 0.65 for GPT-5. It answered in about 425 ms, against roughly 7 seconds for GPT-5. - That contradicted my hypothesis. The gap between 20B and 120B wasn't statistically significant (p = 0.51), so the extra size bought nothing here. - GPT-5 was excellent on SMS spam (0.97 F1) but caught fewer than half of fake news articles. The paper earned a 5 on AP Research. What's next: - A small, fine-tuned open model can be cheaper, faster, and more accurate than a frontier model, but only if you test it domain by domain. I want to keep testing claims like that one. ### SnakeID Point your phone at a snake and find out whether it's venomous, with safety guidance if it is. (https://karthick.me/work/snakeid) - When: 2025 - Role: Builder - Result: Congressional App Challenge district runner-up - Stack: Next.js PWA, AWS Lambda, S3, CloudFront, OpenAI Vision - Live: https://snakeid-karthickaruns-projects.vercel.app - Code: https://github.com/aicoder2009/SnakeID Why I started: - In Arizona, when you run into a snake, the only question that matters is whether it's venomous. I built SnakeID to answer it from a photo, for hikers, homeowners, and emergency responders. What I did: - I designed it first in SwiftUI, then rebuilt it as a Progressive Web App so it installs from any phone's browser. - Your photo goes to an AWS Lambda function that asks OpenAI's vision model for the species, a safety class, and advice, all as JSON. Models don't always return clean JSON, so when the JSON is missing, the function falls back to scanning the text. - The result screen leads with a color-coded safety card and one-tap buttons for 911 and Poison Control. Progress so far: - SnakeID was the district runner-up in the 2025 Congressional App Challenge and earned Special Congressional Recognition from Rep. Andy Biggs (AZ-05). It was my third year entering, after TeenTalk in 2023 and Bills Simplified for Kids in 2024. What's next: - Test it against verified field photos to measure accuracy. ### Khan Academy MCP An open-source MCP server that lets AI assistants search, browse, and read Khan Academy. (https://karthick.me/work/khan-mcp) - When: 2026 - Role: Builder - Result: Open source; install with one command - Stack: TypeScript, Model Context Protocol, npm - Code: https://github.com/aicoder2009/khanacademyMCP Why I started: - Khan Academy had no MCP server, so AI assistants couldn't use one of the best free learning libraries. I built an unofficial one. What I did: - I wrote 12 tools: search, topic trees, courses, lessons, video transcripts, articles, exercises, quizzes, and a study-guide builder. It needs no API key and runs with `npx khanmcp`. Progress so far: - It's published to npm and also runs as a hosted MCP endpoint, so it works with no install. Other developers have starred it on GitHub. What's next: - Wire it into a tutor that builds study plans from real Khan content. ### Infotruster A Chrome extension that uses AI to summarize the Terms of Service and privacy policies people skip. (https://karthick.me/work/infotruster) - When: Summer 2024 - Role: Co-lead, Leangap - Result: Most Outstanding Company; first 10 paying customers - Stack: Chrome extension, OpenAI API, Firebase Why I started: - Nobody reads Terms of Service: they're long and buried in legalese. We wanted people to know what they're agreeing to. What I did: - Our team first planned a full app. I pushed for a Chrome extension instead, so it works right where people browse. - It summarizes long privacy policies, flags data risks, grades how each site handles your data, and suggests safer settings. I worked on the OpenAI API integration and a Firebase cache of results for faster answers. - When our freelance developer fell a week behind, I led the recovery: fixing the website copy and marketing to friends and on social media. Progress so far: - We landed our first 10 paying customers before Demo Day. I gave the winning pitch, and we won Most Outstanding Company. What's next: - The lesson I carried forward: when the plan breaks, find plan B fast. ### Fine-tuning GPT-2 XL on a free GPU My first fine-tune: teaching a 1.5B-parameter model from 2019 to suggest five synonyms or antonyms for any word. (https://karthick.me/work/gpt2-xl) - When: Summer 2025 - Role: Research intern, ASU AI Acceleration Lab - Result: From garbled text to 3–4 sensible words out of 5 - Stack: GPT-2 XL, Hugging Face Transformers, Google Colab T4 Why I started: - I needed to learn fine-tuning, and I wanted to do it on something real instead of following a tutorial. So I made up my own use case: complex enough that prompting alone couldn't solve it, and simple enough to finish. What I did: - I started on AWS SageMaker, but it expects training data in S3 and my dataset lived on Hugging Face, so I switched to Colab's free NVIDIA T4. - My first overnight run (about 12 hours) died at Colab's session limit, so I learned to work within it. Across three runs I tuned epochs, batch size, and learning rate. Progress so far: - Trial 1 produced garbled text. By Trial 3, the model reliably gave five words in a clean format, and 3–4 of them made sense. The biggest jump came from going from 1 to 5 epochs. What's next: - Next time: Mistral 7B or Llama 3 8B with Unsloth QLoRA. 1.5B is too small for good output. ### KTutor An AI tutor that adapts to how each student learns, through video, games, or audio. (https://karthick.me/work/ktutor) - When: 2024 – 2025 - Role: Founder; internship program manager - Result: Two provisional patent applications; ran a 5-intern program - Stack: Generative AI, 3D modeling (Blender), Figma, full-stack web Why I started: - Kids Cloud Club showed me that kids are eager to learn but one human teacher doesn't scale. An AI tutor could reach far more students. What I did: - I designed the AI system and the platform architecture. I'm the named inventor on two provisional patent applications: "System and Method for Delivering Personalized Learning Content" and "System and Method for Generating Learning Content." - The summer after 10th grade, in partnership with the El Segundo School District, I ran an internship program. I managed 5 high school interns and taught them 3D modeling, full-stack development, and UI/UX while they helped build KTutor. Progress so far: - KTutor's lessons fed directly into Lucky, which applies the same idea to accessibility. ### Library of College An evidence bank for college essays: save your stories all year, then let AI find which ones answer each prompt. (https://karthick.me/work/library-of-college) - When: Oct 2026 – now - Role: Builder - Result: Live in private beta - Stack: HTML and JavaScript, AWS Lambda, DynamoDB, Clerk, OpenAI - Live: https://liboc.vercel.app - Code: https://github.com/aicoder2009/library-of-college Why I started: - Supplemental essays reward specific, lived details, and those details never show up on demand when you sit down to write. I hit this in the middle of my own application season. What I did: - I split the problem in two. Capture is a quick dump into seven categories that match how applications are structured, and AI writes the titles and sorts each entry. - Retrieval works by pasting in a real prompt: the model returns your best-fitting stories, labeled strong or possible, each with a concrete reason. - The whole app is one HTML file and one AWS Lambda. Sync is careful: stale writes are rejected and merged by entry, with daily snapshots, so nothing gets lost across devices. Progress so far: - It's live at liboc.vercel.app with a waitlist, and the code is public so others can self-host it. What's next: - Open it to more students this application season. ### OpenCitation A free, open-source citation manager with no ads, no tracking, and no account required. (https://karthick.me/work/opencitation) - When: 2026 - Role: Builder - Result: Web, Chrome extension, and desktop; 668 tests - Stack: Next.js 16, React 19, Tailwind, Clerk, DynamoDB, Electron - Live: https://opencitation.vercel.app - Code: https://github.com/aicoder2009/opencitation Why I started: - Every citation site I used for school was buried in ads and trackers. Citing a source shouldn't cost your attention or your data. What I did: - It generates citations from a URL, DOI, ISBN, or manual entry in APA, MLA, Chicago, and Harvard, plus journal styles through the Citation Style Language. - You can save citations to lists, group lists into projects, share them with a link, and export to BibTeX or RIS. - It ships as a web app, an installable PWA, a Chrome extension, and an Electron desktop app, backed by 668 tests. Progress so far: - It's live and MIT-licensed, and its counter shows 70 citations generated so far. ### Bills Simplified for Kids Upload a bill from Congress and ask it questions, with answers explained so kids can understand them. (https://karthick.me/work/bills-simplified) - When: 2024 - Role: Builder - Result: 2024 Congressional App Challenge entry - Stack: Generative AI, PDF question answering Why I started: - Laws affect kids too, but bills are written in dense legal language that almost no kid will read. I wanted any student to be able to find out what a bill actually does. What I did: - I built a web app where you upload a bill as a PDF, then ask questions about it in plain language. In my demo, I uploaded the Path to Independence Act of 2024. Progress so far: - I submitted it to the 2024 Congressional App Challenge with a recorded demo. A year later, I entered again with SnakeID, which placed district runner-up. ### TeenTalk An AI listener for teens: type what's on your mind and get a supportive, spoken response, with crisis resources when it matters. (https://karthick.me/work/teentalk) - When: Oct 2023 - Role: Builder - Result: My first Congressional App Challenge entry, in 9th grade - Stack: Python, Streamlit, OpenAI GPT-3.5, gTTS - Code: https://github.com/aicoder2009/teentalk Why I started: - Teens don't always have a trusted adult to talk to when something is wrong. I wanted something that would listen at any hour and answer with compassion. What I did: - I built a Streamlit web app where a teen types their concerns. GPT-3.5, prompted to respond like a compassionate school counselor, replies in an encouraging tone, and the reply is also read aloud with text-to-speech. - If a message mentions self-harm or suicide, the app points the user to the 988 crisis line, and a clear disclaimer says it is not a substitute for professional help. Progress so far: - This was my 2023 Congressional App Challenge entry. It was my first AI app, and it started a thread that runs through Lucky today: using AI to reach people who would otherwise go without help. ### AWS Breaks Brakes for an AWS account: pause every running service to stop the bill, without deleting anything. (https://karthick.me/work/aws-breaks) - When: Jan 2026 - Role: Builder - Result: Open-source CLI, installable with pip - Stack: Python, AWS (EC2, RDS, ECS, Lambda), CloudFormation - Code: https://github.com/aicoder2009/awsbreak Why I started: - When an AWS bill runs away, the usual fix, aws-nuke, deletes everything. I wanted an emergency brake that pauses spending and lets you resume exactly where you were. What I did: - I built a command-line tool: run `aws hit breaks` to stop EC2 instances and RDS databases, scale ECS services to zero, suspend Auto Scaling groups, and remove Lambda provisioned concurrency. Run it with `--resume` to put everything back. - It shows estimated monthly savings, has a dry-run mode to preview changes, and runs under a dedicated IAM role with minimal permissions, set up by a CloudFormation template. Progress so far: - It's open source under the MIT license and installs with `pip install aws-hit-breaks`. ### Kids Cloud Club A free online club where I taught other kids cloud computing and AWS over Zoom. (https://karthick.me/work/kids-cloud-club) - When: 2020 – 2021 - Role: Founder - Result: Free classes for kids around the world, including India - Stack: Amazon Web Services, Zoom Why I started: - After I passed the AWS exam at 9, my dad asked what was next. I figured that if I could pass a test many adults find hard, every kid should get access to the same knowledge. What I did: - I started a free club over Zoom and taught cloud computing and AWS to kids, including students in India. - Then I trained my friends to teach, so we could reach more kids. Progress so far: - In 2021 I ran KidCon, Kids Cloud Club's own conference for kids. - But one live class at a time couldn't scale. That limit is what pointed me to AI tutoring, and to KTutor and Setu Labs. ### AWS Robotics internship A summer with the AWS Robotics team, helping test and program their TurtleBot robots. (https://karthick.me/work/aws-robotics) - When: Summer 2019 - Role: Intern, AWS Robotics - Result: Programmed the TurtleBot at age 10 - Stack: Amazon Web Services, TurtleBot Why I started: - At re:Invent, I asked Andy Jassy for a summer internship, and AWS invited me to a summer robotics program. I'd been building robots since FIRST LEGO League. What I did: - I helped the AWS Robotics team with their new robots, the Waffle and the TurtleBot, and got to experiment with the TurtleBot and program it myself. Progress so far: - It was my first look at how a real engineering team builds, and it's where I first saw cloud and robotics come together. ## My story: I was 9 when AWS told me I was too young. This is what I did about it, and how it took me from a website for my dad to building AI for blind and low-vision people. ### Age 8: It started as a surprise for my dad When my dad passed his AWS certification, I wanted to do something bigger than just saying congratulations. So I pulled up GeeksforGeeks, taught myself HTML and CSS, added some JavaScript, and surprised him with a website celebrating his pass. He couldn't believe it. Looking at what I'd built, he asked whether I wanted to try cloud computing myself. I didn't just want to try it—I wanted the certification too. I promised him right then that I'd study every single day for a year, even if on the busiest days it was only for five minutes. ### Age 9: Turned away at the door A year later, we drove to the testing center with my notes in hand. When we walked in, the proctor didn't even look at me—she assumed my dad was the one taking the exam. > "Sorry sir, we don't babysit." (The proctor, to my dad) My dad told her the baby was the one taking the test, and she'd be the one sitting him. It didn't help. AWS only allowed candidates 18 and older, so after 365 straight days of studying, we were turned away in the lobby without me ever touching a keyboard. I could have waited nine years. Instead, my dad and I sat down that evening and emailed Andy Jassy, then the CEO of AWS. We didn't expect someone running a company that size to read an email from a 9-year-old, let alone reply. He replied. The age limit was about liability, he explained, but AWS would make an exception for me. Following that conversation, AWS updated its policy and dropped the age limit from its certification exams altogether. ### July 30, 2018: "Great, Dad, I know I passed" On July 30, 2018, I finally got to sit in front of the testing monitor in Carson, California. When the door opened and I walked back out to the waiting room, my dad looked up anxiously to ask how it went. I told him I already knew. The result reached us an hour later over lunch at the Carl's Jr. next door, when his phone buzzed with the score report. I'd passed. At 9, I was the youngest person in the world to earn an AWS certification. ### 2018 to 2019: If you never ask, the answer is always no That winter, Andy Jassy invited my family to re:Invent, AWS's biggest conference. When he came over to talk on the expo floor, I knew I had one shot to take a leap. I looked up and asked him for a summer internship. > "Karthick, if you never ask, the answer will always be no." (Andy Jassy, then CEO of AWS) That sentence became my operating principle. The next summer, at 10, I worked with the AWS Robotics team, helping test and program their TurtleBot robots. That June I also stood on the keynote stage at the AWS Public Sector Summit in Washington, D.C., speaking to an auditorium of enterprise leaders, and NBC4 Washington covered the story. I told the reporter I was working toward building a robot dog—mostly because my mom wouldn't let me have a real one. ### 2020 to 2021: "So what's next?" Right after my exam results came in, my dad asked me a simple question: "So what's next?" My answer: if a 9-year-old could pass a test plenty of adults struggle with, every kid deserved a shot at learning this. In 2020, as school went remote, I started Kids Cloud Club over Zoom, teaching cloud computing to kids across the country and in India. To reach more students, I trained friends to teach sessions, and in 2021 we organized KidCon, our own cloud conference for kids. Yet one live class at a time could only reach so many students. That ceiling is what first pushed me toward AI: an intelligent tutor that could adapt to each student and reach far more people than I ever could on my own. ### Setu Labs, 2024: Turning a club into a company In 9th grade I built my first AI app, TeenTalk, an AI listener for teens. In 2024, I turned what I'd learned from Kids Cloud Club into a company, Setu Labs, to give students the kind of personalized guidance that used to require a private human tutor. We started with KTutor, an AI tutor that adapts to how each student learns. I designed the AI system, filed two provisional patent applications as named inventor, and led five high school interns who helped build and test it. Later that summer at Leangap, execution hit a real wall. Our freelance developer fell a week behind on Infotruster, our Chrome extension, and Demo Day was less than a week away with no time left to prove traction. Instead of giving up on sales, I got the team rewriting our messaging and reaching out directly to friends and on social media. By Demo Day we had 10 paying customers. I gave our pitch to the judges, and we won Most Outstanding Company. It was an early lesson in startup reality: when plan A breaks, drop the excuses, rally the team, and find plan B before the clock runs out. ### Lucky, 2024 to now: When someone in my family lost their vision While building AI tools, a personal crisis changed everything. A close family member began losing their vision. The reality hit me one afternoon when I watched them struggle through their email: their screen reader kept stumbling over formatting, and they finally had to ask someone else in the room to read private messages to them. Nobody should need another person to read their own email. I started looking at what screen readers actually do. They read every tag and line number in the same flat, robotic voice. Scanned or untagged PDFs often don't read at all, and there is no way to pause and ask a clarifying question about what you just heard. I pivoted Setu Labs to fix this. I built Lucky, a voice-first digital guide dog, named after the real dog my parents finally let me get. You hold Space and talk naturally: it navigates Gmail by voice, reconstructs PDFs in the right reading order, and searches the web with cited sources. I took an early build to a Foundation Fighting Blindness chapter meeting and demoed it for about 300 blind and low-vision users. I asked for direct, unvarnished feedback, and they didn't hold back. They cared about latency, natural cadence, and being able to adjust reading speed instantly. Today, we build Lucky directly with their chapters. At VISIONS 2026, our first national conference booth, I expected questions about technical features. > "People asked about cost instead." (Visitors at VISIONS 2026) We had priced Lucky at $20 a month, like most software. But talking to dozens of visitors made the economic reality sink in: many visually impaired people live on fixed incomes and rely on state vocational rehabilitation agencies or assistive grants for their tools. Lucky worked, but our distribution model was completely disconnected from the people who needed it most. ### Now: The problem I haven't solved yet This journey started with an eight-year-old building a website for his dad and refusing to walk away from a closed door. Today, the mission is much bigger: ensuring blind and low-vision people have equal, dignified access to the digital world. That takes more than a product that works. I want to take Lucky from conference booth demos to daily use, expanding pilots across Foundation Fighting Blindness chapters and figuring out how to reach people who can't pay $20 a month through state programs and grants. Solving that distribution challenge, and learning how to build an enduring company around it, is what comes next. ## Moments ### Commended by the U.S. Senate October 18, 2018: Sen. Dianne Feinstein of California commended me as the youngest person to earn the AWS Certified Cloud Practitioner certification. (https://karthick.me/moments/us-senate-commendation) #### October 2018: In her words Less than three months after I passed AWS's Cloud Practitioner exam, I received a Certificate of Commendation from the United States Senate, signed by Sen. Feinstein. It says why: > "I am proud to commend you on being the youngest person to receive this certification, and on your many accomplishments." (Sen. Dianne Feinstein, on the certificate) #### Age 9: It almost didn't happen AWS only let people 18 and older take its exams, so the testing center turned me away at 9. My dad and I emailed Andy Jassy, then AWS's CEO, and AWS made an exception. I passed on July 30, 2018, in Carson, California. #### 2018 to 2019: One of three honors The Senate's certificate wasn't the first. The California State Senate's Certificate of Recognition, from Sen. Ben Allen, is dated August 24, 2018. Then in March 2019, the Redondo Beach Board of Education honored me at a board meeting, and that night I held all three certificates for a photo with my elementary school principal. ### A year after AWS turned me away, I was on its keynote stage June 11, 2019, Washington, D.C.: At 10, I was the final speaker at the AWS Public Sector Summit keynote, telling the audience how AWS's age limit had turned me away. (https://karthick.me/moments/aws-public-sector-summit-2019) #### Backstage: Getting ready to go on Before going on, I got mic'd up backstage and a high five in the green room. Then the host introduced me as the youngest person ever to become an AWS Certified Cloud Practitioner, and my dad and I walked onto the stage together. #### On stage: "A vast universe" The host's first question was the one they always ask: what's cloud computing? At 10, my answer was that it's "a vast universe" where your data is stored quickly, it's fast and secure, and you can be elastic and flexible with your servers. Then I told the audience how it started. My dad had just come back from taking an AWS certification exam, and I'd built him a website on Amazon S3 with a button that said "click here." When he clicked it, a pop-up said, "God bless you, Dad, I knew you could do it." He asked if I wanted to study for the certification too. I gave him an energetic yes. #### A year earlier: "Nine more years" Then the host asked what happened at the testing center. On my first try, it turned us away, because AWS only let adults take its exams. My dad was disappointed on the way back. I wasn't: > "I told my dad I would come back here in nine more years and take the exact same test." (On stage, at 10) The host told the audience the other half: our email had reached Andy Jassy, and AWS had since changed its policy so that people of any age can take its exams, with a parent's supervision. #### What's next: Three answers Asked what was next, I had three answers: turn the AWS DeepRacer that Andy Jassy gave me into a robot dog, take the developer exam like my dad, and run the AWS coding club I'd started at my school with my principal's approval. The host asked to help with the club, and later added, "If I could hire you today, I would." That day, I was also interviewed on camera on the expo floor. That summer, I joined AWS's summer immersion program with its RoboMaker team, helping test and program TurtleBot robots. ### From no Spanish to the Spanish Honor Society 2025 to 2026, Basha High School: I started Spanish 1 knowing none of it. By junior year, I was a student representative in Basha's Spanish Honor Society. (https://karthick.me/moments/spanish-honor-society) #### Spanish 1: Starting from zero I took Spanish 1 as a freshman with no background in the language, and I wasn't sure I could learn one at all. Words came fast. Conjugations didn't: I'd get the ending right on a worksheet, then slip the moment I had to use it on my own. So I put my strengths to work on my weak spot. Vocabulary and partner conversations turned into grammar practice until the patterns stuck, and because words came easily to me, I helped classmates who got stuck on one. #### Spanish 2: From hardest to best In Spanish 1, presentations were my hardest assignments. Standing up to present in a language I hadn't mastered yet felt harder than any talk I'd given in English, so they got more of my preparation than anything else in the course: I scripted each one and rehearsed it aloud until I could give it without notes. When Spanish 2 ended, they were my strongest work. #### Spanish 3: Why it stuck As a junior, I moved up to Honors Spanish 3 and became a student representative for Basha's Spanish Honor Society. Spanish was the first thing in high school I learned from nothing instead of building on what I already knew. Being a beginner and sticking with it is a habit I've needed ever since, from fine-tuning language models to starting Setu Labs. Sitting on the outside of conversations I couldn't follow also showed me, in a small way, what the blind and low-vision people Lucky is built for face when digital content isn't made for them. ## When plans broke Where it started: building robots in FIRST LEGO League, 2018. Most of what I've built since began with something going wrong. - 2018: AWS only let adults take its certification exams. I studied 5 to 10 minutes a day for a year. My dad and I emailed Andy Jassy, and AWS made an exception, then dropped the age limit. At 9, I became the youngest person to pass the Cloud Practitioner exam. - 2020: I started Kids Cloud Club to teach cloud computing to kids for free, but teaching over Zoom one class at a time couldn't scale. I turned to AI, which could act as a tutor for far more kids. That became Setu Labs and KTutor, an AI tutor with two provisional patent applications naming me as inventor. - 2024: Building Infotruster at Leangap, a startup program, our freelance developer fell a week behind. That left almost no time to go to market. I led the recovery: we fixed the website copy and marketed to friends and on social media. We landed our first 10 paying customers before Demo Day and won Most Outstanding Company. - 2025: On my AI research internship, SageMaker expected data in S3, and my dataset lived on Hugging Face. I switched to Colab's free GPU and reran training three times. The model went from garbled text to 3–4 sensible words out of 5. - 2026: My AP Research hypothesis was that the bigger model would win. It didn't. I reported what the data showed. A fine-tuned 20B model beat GPT-5 on fraud detection: 0.78 F1 vs. 0.65. AP score: 5. ## Timeline - Jul 2018: Took the AWS Cloud Practitioner exam at 9. Age 9, in Carson, California. AWS made an exception, and a year of daily studying paid off. - Aug 2018: Became the youngest AWS Certified Cloud Practitioner. If I could pass at 9, I figured every kid should get access to the same knowledge. - Oct 2018: Built and tested robots for FIRST LEGO League. Age 9, getting our robot ready for competition. - Oct 2018: Programmed our FIRST LEGO League robot. Writing the code for the competition table. A summer later, I was programming robots at AWS. - Nov 2018: Spoke at a re:Invent session for students. Age 9. The host introduced me, and I walked them through what I'd been building. Press play. - Nov 2018: Demoed my work on the re:Invent expo floor. Showing attendees what I'd built at AWS's biggest conference. - Nov 2018: Invited to re:Invent by AWS CEO Andy Jassy. He asked what I was building: a robot dog named Skittles. Press play. - Mar 2019: Honored by the Redondo Beach Board of Education. For being the world's youngest AWS certified person. - Jun 2019: Mic'd up to speak at the AWS Public Sector Summit. Backstage in Washington, D.C., a year after the exam. - Jun 2019: Spoke on the Summit keynote stage at 10. In front of the full Summit audience. Press play. - Jun 2019: Interviewed on the Summit expo floor. Being interviewed on the expo floor, a year after I was told I was too young for the exam. - Jun 2019: High five in the Summit green room. Backstage, at 10 years old. - Jun 2019 to Sep 2019: Summer intern on AWS's robotics team. I asked Andy Jassy for a summer internship, then helped test and program TurtleBots at 10. - Mar 2020 to Jul 2021: Founder, Kids Cloud Club. Teaching AWS for free over Zoom, including to students in India. - 2021: Ran KidCon, a cloud conference for kids. Kids Cloud Club's own conference. But one live class at a time couldn't scale. Press play. - Fall 2023: Founded my school's Hack Club chapter. As a freshman, I ran HTML, CSS, and JavaScript workshops. It later merged into DevOps Club. - Oct 2023: Built TeenTalk, my first AI app. An AI listener for teens, and my first Congressional App Challenge entry. It started the thread to Lucky: using AI to reach people who'd otherwise go without help. - Jun 2024 to Jul 2024: Led Infotruster to Most Outstanding at Leangap. Our developer fell behind. I led Infotruster's recovery: 10 paying customers and the winning pitch. - 2024: Volunteer, Clarendale of Chandler. Leading conversations and memory activities with residents living with dementia. - Sep 2024 to Present: Founder, Setu Labs. I turned Kids Cloud Club's mission into an AI company, starting with KTutor, an AI tutor. - Oct 2024: Earned AWS Certified AI Practitioner. As an Early Adopter, six years after my first AWS certification. - Oct 2024: Built Bills Simplified for Kids. Laws affect kids too, so I built Bills Simplified for Kids, my second App Challenge entry. - May 2025 to May 2026: Junior Blood Drive Coordinator, NHS. Learned to run our quarterly drives with Vitalant. My co-coordinator and I ran the year's last drive ourselves: 140+ donors. - May 2025 to Jul 2025: AI research intern, ASU AI Acceleration Lab. SageMaker didn't fit my data, so I fine-tuned GPT-2 XL on a free GPU. - Jun 2025: Completed Dale Carnegie High Impact Presentations. Two days of public speaking coaching that I now use to pitch Setu Labs. - Summer 2025 to Dec 2025: UN Global Youth Excellence and Leadership program. Made a policy video on education and digital access. Certified in Part 1 that December. - Summer 2025: Ran a 5-intern program for KTutor. With the El Segundo School District, I taught 3D modeling, full-stack, and UI/UX. - Sep 2025 to May 2026: Member, Gilbert Mayor's Youth Advisory Council. One of 30 students chosen to advise the mayor. - Fall 2025 to 2026: Student representative, Spanish Honor Society. Representing Basha's chapter. - Dec 2025: SnakeID: App Challenge district runner-up. My third App Challenge entry earned Special Congressional Recognition from Rep. Andy Biggs. - Spring 2026: AP Research: my fine-tuned 20B model beat GPT-5. My hypothesis was that the bigger model would win. I reported what the data showed, and the paper earned a 5. - May 2026 to Present: Senior Blood Drive Coordinator, NHS. Now training the junior coordinators and running drives with our new partner, the American Red Cross. 330+ donations to date. - May 2026: Completed my Mayor's Youth Advisory Council term. Helped plan Student Government Day and was chief of emerging technology at the mock council. - 2026: Earned the AP Capstone Diploma and 4.0 Award. Also AP Scholar with Honor and College Board National Recognition. - Jun 2026: Elected Precinct Committeeman at Boys State. At Arizona Boys State, I also co-founded East Valley Law. - Jun 2026: Earned an Arizona CTE certificate in Software & App Design. From the Arizona Department of Education. - Jun 2026: Exhibited Lucky at VISIONS 2026. Demoing Lucky, my voice AI for blind people, at the Foundation Fighting Blindness conference. - Jun 2026: Demoed Lucky to a full booth. Visitors stopping to see what I built after a family member lost their vision. - Jun 2026: Co-developing Lucky with Foundation Fighting Blindness. Its chapters help shape Lucky and test it. - Fall 2026 to Present: President, Basha DevOps Coding Club. I run Day of Code for middle schoolers and our outreach events. ## Recognition - Youngest AWS Certified Cloud Practitioner: Youngest in the world, at age 9 (2018); featured on NBC4 Washington - U.S. Senate Certificate of Commendation: From Sen. Dianne Feinstein, for being the youngest person to earn the AWS certification (October 2018) - California State Senate Certificate of Recognition: From Sen. Ben Allen, as the youngest person in the world to earn the AWS certification (2018) - Honored by the Redondo Beach Board of Education: For being the youngest person in the world to earn an AWS cloud certification (March 2019) - Speaker, AWS Public Sector Summit: On the 2019 keynote stage, at age 10 - Congressional App Challenge: Entered three years running. SnakeID (2025): district runner-up and Special Congressional Recognition from Rep. Andy Biggs - Most Outstanding Company, Leangap: Infotruster; gave the winning Demo Day pitch (2024) - Two provisional patent applications: Named inventor, personalized AI learning content (KTutor) - AWS Certified AI Practitioner: Early Adopter (2024) - Dale Carnegie High Impact Presentations: Completed the two-day public speaking seminar (2025) - GYEL Certified, Part 1: Global Youth Excellence & Leadership Program (Dec 2025) - AP Capstone Diploma: AP Scholar with Honor, 2025 and 2026 - Arizona CTE Technical Skills Certificate: Arizona Department of Education, Software and App Design (2026) - College Board National Recognition Program: School Award - 4.0 Award: Basha High School, for outstanding academic achievement ## Leadership - President, Basha DevOps Coding Club. Co-organizes Day of Code for middle schoolers and the club's community outreach. Founded the school's Hack Club chapter as a freshman. - Blood Drive Coordinator, National Honor Society: junior coordinator, now senior. Four drives a year, 330+ donations. - Arizona Boys State. Elected Precinct Committeeman; co-founded East Valley Law. - Gilbert Mayor's Youth Advisory Council. One of 30 selected to advise the mayor. ## Next What I want to build and learn during the Academy year. - Scale Lucky: Lucky works in demos. Next, it has to work every day, for real people. I want to go from booth demos to daily users in the blind and low-vision community, starting with the Foundation Fighting Blindness chapters that already test it, and find out what keeps someone using it on day 30. VISIONS also taught me that $20 a month is the wrong price: many of the people Lucky is for live on fixed incomes and rely on state agencies to buy their tools, so I'll work out grants and state programs to reach them. - Learn to build a company: I can build products and pitch them. Next, I want to learn how to raise money, hire people better than me, and build a team that wins. Those are things I can't teach myself from a repo. - Work inside an AI lab: I've fine-tuned models and built products on top of them. I want to spend my co-op at Anthropic — to understand how frontier models handle edge cases and uncertainty at the alignment level, and bring that rigor back to Lucky. When a blind user's voice assistant mishears a command or hallucinates a document summary, the stakes are real. ## Press - [NBC4 Washington: Meet the 10-Year-Old Helping Amazon Web Services's Robotics Team](https://www.nbcwashington.com/news/local/meet-the-10-year-old-amazon-cloud-expert/153863/), June 12, 2019. NBC4 covered how AWS turned me away at 9, then dropped its age limit so I could take the Cloud Practitioner exam, and the AWS summer robotics program it led to. ## Links - Site: https://karthick.me - GitHub: https://github.com/aicoder2009 - LinkedIn: https://www.linkedin.com/in/karthickarun