I build products to solve problems wherever I find them.

CET Sandbox is Israel's largest hub for defense-tech startups looking to commercialize in the United States. Founded in 2025 as a direct response to the U.S.-Israel Joint Strategic Dialogue on Critical and Emerging Technologies — and accelerated by the wave of innovation in Israel's defense sector following October 7 — the organization bridges Israeli dual-use startups in AI, autonomy, and quantum with U.S. Department of Defense and Defense Innovation Unit solicitations, while also supporting entry into private U.S. markets. It operates through deep relationships inside the U.S. government and years of expertise navigating the Israeli defense-tech ecosystem.

I joined as a Solutions Engineer intern, bringing fluency in both U.S. business strategy and the Israeli startup landscape. My job was to consult with the startups directly, helping them orient their products toward U.S. government solicitations, sharpen their positioning, develop market entry strategies, and connect with defense-aligned investors in the U.S.

Doing this job well required holding an enormous amount of information simultaneously — deep knowledge of every startup in our pipeline, mastery of active government solicitations and their specific requirements, and awareness of the proprietary intelligence we had from partner meetings and classified briefings. Our team was highly diverse: former military commanders, ex-government advisors, VC investors, bankers, and technologists, from multiple countries and cultural backgrounds. That diversity was a strength in the room. But when people went their separate ways, the knowledge didn't travel with them. Everything lived in documents, in people's heads, and in institutional memory that was hard to access and impossible to take full advantage of.

I was an intern. I knew I'd be leaving and I didn't want the work to stop when I did.

I built a multi-layer internal RAG system — a retrieval-augmented generation tool that lets you ask questions and get answers grounded in your actual documents. I used three core components: Ollama to run the language model locally, keeping sensitive data off external servers; LlamaIndex to structure and index our documents so the model could retrieve the right information at query time; and FAISS for fast vector search across large volumes of unstructured text. The system had three layers: an open-source scraping layer for new startups entering the pipeline, a proprietary layer that processed our internal and classified startup documents, and a U.S. government layer that synthesized active solicitations with notes from our government meetings. Most often, the output was GTM strategy documents and USG alignment recommendations generated for each startup on demand, in addition to quick look-ups, like “Show me all startups in drones we engage with over $10M raised”. I chose this stack over a ChatGPT-based integration deliberately, because Ollama gave me the customizability and data control the sensitivity of our work required.

The original system has since been replaced by a newer tool with more advanced integrations. But the core idea and the underlying principles are still in use. I built it knowing I was leaving, which meant building it for people who weren't me, and that constraint made it better.

The summer between my freshman and sophomore years of college, I interned at TikTok in Business Development. I met with Fortune 500 clients in meetings, pitching them on our partnerships. I briefed internal Product and Engineering teams on what product specs would look like, and I helped write and deliver our global distribution strategy to teams and offices around the world on product partnerships.

I learned so much in the role, but I was inspired by others in the company, who frequently were intrapreneurs at TikTok to do more than just my 'job'. I started developing my own product ideas on the side — thinking through features and partnerships that I thought TikTok was missing. One of them caught my manager's attention. He encouraged me to take it further, and I did.

TikTok has publicly stated its ambition to become a more powerful search platform, moving beyond entertainment to serve users who come with intent. I had access to internal data on TikTok's highest-volume searches and user behavior patterns, and I cross-referenced those against Google Trends data to find where the two overlapped and where TikTok was underleveraging demand. I found a specific user vertical that was actively searching on TikTok but not being served well, and researched and developed a solution.

I built a 20-slide research deck from scratch on internal data, trend analysis, user behavior, and a clear articulation of the opportunity. My research indicated that adding the feature could likely increase engagement and session length for the target user segment. I designed the mockups myself in Canva. And because I came from a product partnerships background, I did the initial BD work in parallel, thinking through how a partnership or integration would support the feature from day one. The pitch was self-initiated. I got in front of the lead PM in TikTok Search, received feedback, iterated on the deck, and had follow-up meetings with people inside the organization who directly oversaw the user segment I was building for.

I left at the end of my internship but stayed in contact with my team, who confirmed it made it to TikTok's deal pipeline in early 2025. There's been a lot of change at TikTok since then, and I don't know its exact status today. But I'm proud that something I built as a freshman in college — on my own initiative, outside the scope of my actual job — outlasted me at the company. It's the clearest signal I have that thinking like a builder, not just an operator, is something I can do anywhere.

Course Vault is a web application that converts course material into an interactive, hierarchical database for custom learning. You upload a syllabus, and the system breaks your course down into chapters, topics, and concepts — each enriched by AI with a summary, vocabulary, and examples.

The inspiration came from object-oriented programming, or OOP — a way of writing code where you organize everything into objects, which are individual units that hold specific properties and belong to larger categories called classes. When I first learned OOP, I saw a direct parallel to how knowledge is structured in any course: a concept (like "Structuralism") has properties (a definition, examples, vocabulary), belongs to a topic (Schools of Thought), which belongs to a chapter (History of Psychology), which belongs to a course. The hierarchy felt obvious once I saw it.

What surprised me was how much better I performed in my liberal arts courses when I started organizing them the same way my CS courses were naturally organized. I did some research and found the science backed it up: we encode new information more effectively when it attaches to an existing structure. Schema theory explains why — a framework helps your brain interpret incoming information instead of treating it as unrelated bits. Hierarchical organization also reduces cognitive load by letting you chunk related ideas together, protecting limited working memory and making retrieval significantly stronger. I wanted to give that structure to any course, and use AI to fill the gaps quickly.

Most study tools — Quizlet, Anki, Notion — are flat. Flashcards are a list. Notes are a document. They give you content but not architecture. You have to hold the structure in your head yourself, which is exactly the cognitive load problem they're supposed to solve.

The deeper opportunity I saw was that the meticulously organized structure that makes CS intuitive — where concepts are nested inside topics, inside chapters, inside courses — almost never gets applied to liberal arts. History, psychology, economics: these subjects have just as much internal logic, but students rarely engage with them that way. I wanted to change that.

I built the application using Claude and Cursor as my primary development tools, working through the architecture iteratively. Before writing any code, I ran informal conversations with friends about what the right UI should look like — what made sense, what felt confusing, what they'd actually use.

The biggest design tension was around the course map view. One camp felt that showing all concepts across all topics at once was overwhelming — too much on screen. Another felt that seeing the full hierarchy simultaneously was the whole point, the thing that made the product different from everything else. I tried to navigate both sides through a staged workshop flow: you build the structure level by level (chapters, then topics, then concepts), approving each layer before moving to the next. That way the map feels earned by the time you see it in full.

On the technical side, the hardest problems were the API integrations. Claude's responses kept getting cut off mid-JSON when token limits were too low, and internal fetch calls between Next.js routes were failing silently. I learned to debug by adding structured console logs at each step of the pipeline, reading the raw API responses before parsing, and isolating exactly where in the chain things were breaking. Every bug had a specific cause — and finding it was a matter of narrowing the problem systematically until I could see it clearly.

Course Vault is live and working. I've found that it performs better on science-heavy courses — psychology, biology, introductory CS — where syllabi tend to be structured and concept boundaries are clear. Upper-level liberal arts courses, where a single "concept" might span decades of historiography, are harder to break down cleanly, and that's the next problem I want to solve.

The feature I'm most excited to build next is an auto-upload flow: you drop in a single page of notes or a slide deck, and the system figures out exactly where in your course map that content belongs — no manual placement required. The goal is to make adding material feel effortless enough that students actually do it.

The point is to eventually make the uploaded content from a course the primary content in the database, and the AI-generated content secondary.

Sitting in the back of AP Environmental Science, having submitted our college applications, my classmate Ethan Schulman and I could not stop talking about business ideas. Reflecting on our high school experience, our ideas came from problems in markets we understood — quickly leading to Tutor Tree.

We spoke about what it actually takes to ace a course. While Mathnasium and Kumon were charging $100–150 a session, the students getting the best grades weren't coming from those places — they were getting help from students who previously had the same teacher, in the same course, and knew exactly how to succeed. Simultaneously, there were top students at our high school, who didn't realize they could put their knowledge to use as tutors, and others with only one consistent student to tutor.

We built Tutor Tree to solve both of these problems for students at our high school in a centralized, peer-to-peer tutoring marketplace.

Highlights: In two years of operation, we generated roughly $20K in revenue, facilitated over 300 tutoring sessions, brought on 100+ student-tutors, and expanded to four New York high schools. Ethan and I closed it the summer before our junior years of college but constantly reflect on our learnings and memories.

The tutoring market at the high school level was broken in a specific way: it was expensive, generic, and mismatched. Professional tutoring companies charged premium rates for adult tutors who hadn't taken AP Bio or Regents Chemistry in decades. Meanwhile, students were already tutoring each other informally — through group chats, after-school arrangements, word of mouth — but with no infrastructure, no trust signals, and no way to scale.

The insight was that the best possible tutor for a 10th grader struggling in AP US History is a junior who aced the same class with the same teacher last year. They know the specific essay prompts. They know what gets marked down. They speak the same language. We just needed to build the platform that connected them, and gave both sides a reason to show up.

Building a two-sided marketplace means you have two products to think about simultaneously: one for students, one for tutors. To create hype around Tutor Tree, we made commercials, collaborated with local businesses, created company merchandise, and more. Each school operated through a Tutor Tree ambassador that handled the management of tutors, ensuring they showed up to sessions, were briefed as necessary, etc. Quickly, we spread through word of mouth for having brilliant tutors, exceptional prices, and a young entrepreneurial feel.

On the product side, we aimed high trying to have a superior platform to major players. We did so through a complex pipeline of integrations: Stripe for payments, Square for bookings, and JotForm plus email for personalized matching. We added multiway booking, in-site messaging, automated payment splitting. We established a sample group of parents, students, and tutors to give us feedback on UI/UX.

There were many operational challenges as well. A tutor wouldn't show up and we'd have to recruit a replacement in hours, or step in ourselves. An angry parent would call upset about a session. Tutors and students, once paired, would sometimes take the relationship off-platform — so we had to think carefully about what kept people coming back. Managing a distributed team of over 100 high school students across multiple schools, while being college freshmen ourselves was hard. It taught me more about people and operations than anything I've studied since.

The site is no longer live. We took it down, split the earnings, and moved on. But the value we created didn't disappear when the platform did: Students got tutoring, high school kids made money, and mini communities formed inside schools that wouldn't have existed otherwise.

I am most proud of the impact we had on the students in the years below us. I still get messages from former tutors and students about business ideas they're working on. At 21, getting to play any role in inspiring someone to chase something entrepreneurial is something I am extremely grateful for. Tutor Tree was my first real proof that you could build something, make it work, and have it mean something to the people it touched.