Discover
Startup Project: Build the future
Startup Project: Build the future
Author: Nataraj
Subscribed: 4Played: 24Subscribe
Share
© Nataraj
Description
Conversations with founders, operators and investors who are building the future. Listen to find the stories, ideas, tactics & investments behind the products that will define the future of technology.
https://startupproject.substack.com/
https://startupproject.substack.com/
130 Episodes
Reverse
Sid Sheth, co-founder and CEO of D-Matrix, joins Nataraj on Startup Project to talk about the future of AI inference, why GPUs may not be the answer for every workload, and what it takes to build a semiconductor startup.Key takeawaysTraining and inference are like education and a career. Sid's view is that inference is the much larger opportunity.D-Matrix focuses on ultra-low-latency inference using SRAM and in-memory compute, not on throughput.Customers now deploy racks and clusters, so D-Matrix moved from chip scale to rack scale.Raptor stacks custom DRAM on compute. Sid says this gives roughly 20–30x more capacity per rack.Timestamps0:00 – Intro: from training to low-latency inference0:43 – Why start a chip company in 2019, before ChatGPT4:04 – Betting on inference for the data center and cloud7:08 – Transformers, GPT-3 and finding product-market fit9:16 – The first chip and the in-memory compute demo14:25 – How a first chip gets built: partners, TSMC and COVID17:10 – From a digital in-memory engine to a first product in 202420:24 – What "rack scale" means and why the industry moved to it24:01 – D-Matrix vs. GPUs: throughput vs. interactivity, and no HBM26:58 – The business model and target customers29:07 – The Nvidia partnership and the stacked-DRAM product, Raptor34:32 – The two acquisitions and how they fit37:06 – The AI hardware investment cycle40:05 – What success looks like for D-Matrix
Nataraj speaks with Mike Collins, founder of Alumni Ventures, about how his firm opened up venture capital to individual investors through pooled funds and co-investing. They also dig into the structural problems in private markets, why companies stay private longer, and how AI is reshaping venture without eliminating the need for patience, judgment, and diversification.Key topicsMike Collins’ background in ventureWhat Alumni Ventures is built to doHow the alumni fund model worksWhy diversification matters in ventureThe long-term nature of venture investingProblems in today’s venture and public marketsWhy companies stay private longerAI, hype, and real innovationWhy adoption is slower than demosHow new venture firms get startedLearning resources for retail investorsTimestamps00:00 - Introduction and why Alumni Ventures matters00:24 - Mike Collins’ VC background and founding Alumni Ventures01:24 - The mission: access and education for individual investors03:59 - How an alumni fund is structured and invested05:55 - Why pooled capital and network scale matter06:24 - Venture capital as a core engine of the economy07:28 - Why most individuals need private-company exposure08:51 - Why venture portfolios need diversification10:33 - Green D fund leadership and how team sourcing works11:52 - How deals get allocated across multiple funds13:11 - Venture as a long-term, slow-compounding asset class14:07 - Building a venture allocation over several years15:04 - Why checking investments too often looks like trading16:44 - The biggest accessibility problem in venture today17:42 - Why public-company incentives have weakened18:43 - Why companies delay going public20:16 - Late-stage private companies and who captures the upside21:14 - Why SpaceX illustrates the privatization of gains22:51 - Public versus private markets and the role of competition24:19 - AI as a real platform shift, not just hype26:15 - Big opportunities beyond AI: energy, defense tech, healthcare27:41 - The cultural bias toward doom and negative headlines29:37 - Why staying private can help companies like Stripe32:40 - How AI capital concentration affects the broader venture market34:53 - The Series A squeeze and how market corrections happen36:49 - The main paths to starting a new venture fund39:15 - Why great companies still take decades to build40:29 - Why AGI timelines are often faster in theory than in reality42:24 - Regulatory backlash and the slower pace of adoption43:18 - Self-driving cars as a cautionary example for AI timelines44:54 - Why the last 5 percent of product adoption is the hardest46:58 - AI as a tutor for learning venture capital47:53 - Books and frameworks for understanding startups and VC49:18 - How to filter noise and think in decades, not days50:44 - Why the best investors often do the least trading51:50 - Mike’s core investing rules and closing thoughts
Postman is the world's leading API platform, trusted by more than 40 million developers and 98% of the Fortune 500. In this episode, we trace how Postman went from a developer's side project into a global, AI-native enterprise platform. Abhinav Asthana shares how Postman reached profitability before raising venture capital, why developer products are so hard to monetize, why he refused to sell and moved from India to the US, and how AI agents are reshaping the company's next decade.Key topics:Building Postman as a side project to solve API testingReaching Ramen profitability with in-product upgradesWhy developer products are hard to monetizeTurning down acquisition and moving from India to the USSingle-player tool to enterprise API platformBuilding in India vs. the USAI agents as a new class of API consumersAstro, Passport, Fabric, and Fern — the agent-native stackThoughts on "token maxing" and AI spendTimestamps:00:00 - Introduction: Postman and Abhinav's journey02:15 - Before Postman: becoming a developer in India05:06 - BITS 360: the virtual campus tour that started it all07:09 - Monetizing mobile and the rise of Instagram08:32 - Parting ways and hacking on Postman on the side09:08 - The original problem: testing APIs11:05 - From side project to the main thing11:58 - The first signals of demand (Chrome Web Store)13:18 - Reaching Ramen profitability with upgrade packs15:52 - Realizing Postman was a venture-scale business17:46 - The enterprise journey and product validation20:22 - Why he refused to sell Postman22:10 - Moving to the US as an Indian founder24:38 - What made Postman compound27:01 - Why developers are hard to serve28:32 - How vibe coding and agents change API consumption35:02 - Building in India vs. the US38:59 - How AI changed how Postman operates41:22 - Thoughts on "token maxing"43:55 - Inside Astro: an agentic operating system45:17 - The agent identity Passport47:08 - Where Postman goes in three to five years48:27 - IPO plans?Resources & Links:Postman → https://www.postman.com/Abhinav Asthana → https://www.linkedin.com/in/abhinavasthana/Nataraj Sindam → https://www.linkedin.com/in/natarajsindam/Startup Project Episodes → https://thestartupproject.io/episodes
Startup Project sits down with Sviat, CEO of Bright Machines, to unpack how the company is using software-first robotics to manufacture complex electronics closer to where they’re deployed. The conversation focuses on why AI infrastructure is a strategic category, how Bright Machines differs from traditional contract manufacturing, and what onshoring really means for speed, quality, and security.Key Topics:In this episode, Sviat explains that Bright Machines is focused on AI infrastructure, specifically the electronics that go inside modern data centers, including compute nodes, storage, and racks.He traces the company’s thesis back to a broader idea: use software and robotics to manufacture electronics anywhere, then narrow that focus to the data center market as demand became clearer.The discussion breaks down the market stack, from chip designers like NVIDIA and AMD, to ODMs, OEMs, hyperscalers, and contract manufacturers.Sviat shares why data center hardware became the right bet before ChatGPT accelerated the market: the products are expensive, strategically important, and driven by quality and throughput more than labor cost alone.The show compares traditional assembly lines with Bright Machines’ approach, which uses more robotics, sensors, cameras, traceability, and humans in the loop where automation does not make sense.Sviat explains how Bright Machines starts with design, using Bright Designer to simulate and improve manufacturability before lines are built, which helps reduce bottlenecks and improve automation over time.He says the company’s main differentiator is its software platform, which orchestrates the line, powers smart skills for navigation and inspection, collects data, and feeds insights back into design.The conversation covers line flexibility, including how much can be reused when switching between CPU, GPU, or different accelerator-based server designs, and when end-of-arm tooling must change.Sviat says Bright Machines is growing rapidly, expects more than 3x growth this year, and can produce high volumes from a small number of sites because of robotics efficiency.The episode closes on the broader case for onshoring AI infrastructure manufacturing in the US: security, time to market, quality, and a labor shortage that makes robotics necessary.Timestamps:06:39 - The market stack: chip designers, ODMs, OEMs, hyperscalers, and CMs09:07 - Why Foxconn, Jabil, and similar contract manufacturers matter10:04 - Why large factories still rely on massive manual labor12:20 - Why data centers are different from cheap consumer electronics13:49 - Security, strategic sectors, and why AI infrastructure belongs onshore16:26 - The first Bright Machines product: CPU compute servers for a hyperscaler17:58 - How the line works: modular stations, yields, and automation levels19:26 - Bright Designer and design-for-manufacturing feedback loops21:20 - Robots, sensors, traceability, and humans in the loop22:19 - Why time to market matters as much as cost23:31 - Yield and throughput: 98% line-level yields and up to 2x throughput25:25 - The Bright Robotic Cell and how the assembly line is structured27:35 - Reusability across products and when tooling changes are needed30:31 - Manufacturing as a service, not repair or field service31:24 - Growth, gigawatt-scale capacity, and output from a single site33:00 - Why current hyperscaler capex is not expected to slow near term34:45 - The bottlenecks before deployment: chips, components, power, permits36:54 - Bright Machines’ three pillars: platform, data layer, and Bright Designer39:15 - Why humanoid robotics is exciting but not ready for industrial use41:16 - Where LLMs and newer AI tools can help the robotics workflow43:57 - The overlooked advantages of onshoring manufacturing in the US45:59 - What Bright Machines could build next: more complex electronics and future AI devices
Dive into the evolving role of AI in design and collaboration as Danny Wu, Head of AI Products at Canva, shares insights on how the platform is transforming creative workflows, democratizing design, and leveraging large language models and diffusion techniques.In this episode:How Canva redefined abstraction layers in design, moving from pixel edits to object-based workflowsThe evolution of AI at Canva: from traditional ML to transformers and large language modelsThe impact of ChatGPT integration on Canva's user experience and business growthAgentic AI: Canva’s approach to AI that acts as a collaborative partner in designChallenges and misconceptions about AI generative models in creative industriesFuture plans: video content tools and more AI-powered featuresTimestamps:00:00 - Introduction to Canva’s innovation in abstraction layers01:12 - Danny Wu’s background and journey at Canva02:46 - Transition from software engineer to Head of AI Products04:25 - How diffusion and large language models accelerate Canva’s AI capabilities06:50 - The dominance of transformer models in Canva’s AI strategy08:51 - Shift from pixel to object, then conceptual design with AI09:19 - Limitations of chat-based creativity vs direct manipulation11:01 - The future of design involving AI-generated, editable content13:30 - Launch of Canva AI and the platform's new architecture15:22 - Use cases and limitations of AI in visual and video content17:05 - Canva’s diverse user base and how AI personalization fits different needs18:48 - Challenges of AI aesthetics and user customizations20:32 - Amazing AI features like Magic Layers for editable image designs22:16 - Comparing models: diffusion, open source, and proprietary tech27:36 - Exploring agentic AI: Canva's vision of AI as a collaborative partner30:37 - How ChatGPT and similar tools boost Canva’s reach and usability34:26 - Personalizing AI output for users and reducing generated “cookie-cutter” content37:38 - Managing AI's creative style and avoiding homogenization42:39 - Canva’s feature development process and testing workflows46:36 - Surprising use cases, like self-grading quizzes in education48:59 - Overlap and differentiation between Canva and other design tools like Figma50:54 - Future video tools and content creation enhancements at Canva








