AI Builders

Rapidata provides human feedback to AI model companies, focused on taste and soft human opinions rather than factual knowledge. Its annotators are volunteers who opt in to give feedback inside apps instead of watching ads, which puts Rapidata’s tasks in front of people through ad slots in thousands of apps.In a recent episode of BUILDERS, we sat down with Jason Corkill, CEO at Rapidata, to learn how the company grew from $25,000 to $30,000 in revenue last year to three and a half million this year, why it chose to sell into a market of roughly 50 customers, and why it avoids competing as a traditional data annotation company.Topics Discussed:Why traditional data annotation is self-cannibalizing, and how Rapidata positions around itWhy taste and soft human opinions are the feedback models still lackHow Rapidata sources annotators through ad slots instead of building its own appWhy volunteer annotators are easier to quality-filter than paid onesBuilding user trust when people expect every ad to redirect themWhy app publishers can prefer Rapidata’s format over traditional advertisingThe decision to accept customer concentration in a market of about 50 companiesHow researchers become champions and where procurement slows larger dealsWhy open datasets on Hugging Face and research conferences drive Rapidata’s sales motionChoosing a lane when the core technology could serve many markets

Reactor is building the platform infrastructure to serve real-time AI, starting with world models: models that take in inputs, keep track of what came before, and generate what happens next live instead of after a long wait.In a recent episode of BUILDERS, we sat down with Alberto, Taiuti, CEO and co-founder at Reactor, to learn why he sees other companies in his space as evangelists for the category, how Reactor found the infrastructure gap the rest of the market had raced past, and how the team decides which markets to educate first.Topics Discussed:Why Alberto welcomes companies doing similar workWhy the race to build world models left deployment unsolvedHow Reactor spent months trying to disprove its own idea before committing to itHow market understanding of world models went from explaining them in meetings to conferences in about a yearWhy robotics, customer support, and streaming teams understand Reactor immediately while other industries still need educationHow Reactor started in media and entertainment and built one inference engine that serves multiple verticalsWhy solving audio, video, and command sync for customer support avatars also solves robotics problemsThe risk of real-time AI being seen as useful only for robotics

In a recent episode of Unicorn Builders, we sat down with Eric Ho, Co-founder and CEO of Goodfire, to learn how an exploration-driven interpretability lab is turning foundational research into an enterprise product, and how one public AI incident changed the conversation around its category.Topics Discussed:Why Eric left the SaaS company he spent seven years building after ChatGPT launchedWhy even the researchers building frontier models can’t explain how they workWhy Goodfire modeled itself on Bell Labs, Xerox PARC, and early OpenAI instead of a SaaS playbookWhat it means to run a research lab like an “overgrown garden”How the first two years broke into chapters: research previews, recruiting leading interpretability researchers, then sharpening research directions and building SilicoWhat the OpenAI and Hugging Face incident revealed about reward hacking, chained exploits, and agent swarmsWhy Silico sits one level of abstraction above Claude Code and CodexWhy Goodfire sells top-down to enterprises, with contracts one to two orders of magnitude larger than Eric’s SaaS dealsWho buys interpretability: frontier labs, neolabs, data labeling companies, and enterprises training their own models

Toloka is an AI data platform that produces training data, evaluations, and human expert contributions to help AI labs and enterprises build higher quality models. In a recent episode of Front Lines, we sat down with Olga Megorskaya, Founder and CEO of Toloka, to learn how the company has navigated every major phase of AI development and what she actually believes about the future of human expertise in AI systems.Topics Discussed: How Toloka has evolved from basic crowdsourcing to RLHF to RL gyms across a decade of AI developmentWhy the shift from foundational models to post-training and fine-tuning is driving new demand for human dataHow the types of human expertise Toloka requires have shifted as AI capabilities grewWhy managing human efforts at scale is, in Olga's view, a purely technological problemThe customer concentration risk that came with the era of foundational modelsWhy enterprise AI adoption is Toloka's main bet for diversifying beyond frontier labsWhy generating AI output is now easy and why verifying it remains the hard problemHow physical AI and robotics is creating a new demand curve for human dataWhether expert recruitment has gotten harder or easier as AI data work has become more respectedThe statistical methods from 1979 that Toloka's quality management systems are still built on

Building the real-time infrastructure layer for voice AI is a narrow technical bet that turned into a platform play. LiveKit started as an open source project during the pandemic and grew into the stack that powers OpenAI's voice mode for ChatGPT, handling the low-latency network infrastructure that connects users to AI models anywhere in the world.We sat down with Russ d'Sa, co-founder at LiveKit, to learn how an open source project with no commercial intent became the foundational infrastructure for voice AI applications, and what it took to recognize product market fit even when every signal was pointing directly at it.Topics discussed:- Why LiveKit started as an open source project during the pandemic and hit the GitHub top ten repos across all languages within two weeks of launch- The acquisition offer LiveKit received on day one of its first open source release, and why Russ asked for $100 million- How OpenAI signed up for LiveKit in secret using a personal email address, three weeks before reaching out to start a commercial relationship- Why Russ denied product market fit twice despite having a top GitHub repo, enterprise interest from companies like Spotify and Reddit, and an inbound from OpenAI- The difference between stateful and stateless computing, and why the entire software development lifecycle has to be rebuilt for voice AI applications- Why going all in on voice AI took months longer than it should have, and what Russ would do differently- How LiveKit built its full platform stack from network infrastructure to agents framework, evals, inference routing, and observability- Why all of LiveKit's growth to date has been inbound, and what Russ is now building out on the go-to-market side for the first time- How the OpenAI partnership created two separate demand channels: enterprise inbound from the logo halo effect, and developer adoption by association

Cyberhill Partners builds and deploys enterprise AI solutions, including Cerebro, an enterprise AI platform Rob describes as AI in a box, and Wolverine, a digital twin product, drawing on eight years of AI work inside the intelligence community.In a recent episode of BUILDERS, we sat down with ⁠Rob Buller⁠, Chief Executive Officer of ⁠Cyberhill Partners⁠, to learn how the company is bringing intelligence-community-grade AI to enterprise buyers who are still missing the semantic layer that makes AI work.Topics Discussed:Why the core principles of technology adoption have not changed, but complexity has, splitting the buyer across CIOs, CTOs, and CISOsThe AI factory model: a runtime AI fabric that plugs into Snowflake, Databricks, Grok, or Gemini instead of shipping compiled, embedded logicHow Cyberhill positions Cerebro against Palantir on cost, vendor lock-in, and implementation speedWhy Rob has been shocked that heads of AI at multi-billion dollar companies do not know what an ontology isHow large incumbents like Workday and Salesforce try to freeze the market when new technology emergesWhy Cyberhill picks verticals by following demand into government, automotive, and healthcareMarketing at the local level in secondary markets like Dallas and Denver with billboards, airports, and quarterly steak dinnersPartnerships with ServiceNow and Databricks, and government AI work including a global biosurveillance platformGTM & Technology Adoption Lessons:The market cannot buy what it does not understand: Rob has been on roughly a hundred client calls where heads of AI at multi-billion dollar companies could not define an ontology. Without the semantic layer, he argues, you cannot apply context to AI or get traceability. Until buyers understand that, adoption stalls, so education is now part of the sales motion whether Cyberhill wants it or not.Sell the problem solved, not the underlying technology: "People don't care about ontologies and knowledge graphs, they care about can you solve my problem." Cyberhill leads with the business problem and only opens up the technical underpinnings when a buyer wants to know why the product is different.Expect incumbents to freeze the market: Rob says large companies respond to new AI entrants by telling customers they already have AI covered. He admits it does not really work, but it works a little bit. Plan the GTM knowing buyers are hearing "you don't need AI" from vendors they already pay.Let demand pick your verticals: "I've found that business is a lot like water. It finds the lowest level." Rather than forcing a vertical strategy, Cyberhill follows where demand shows up: government, automotive, healthcare. When healthcare demand grew, the company hired a doctor, because subject matter experts are how it enters an industry credibly.Position against the expensive incumbent on speed to value: Rob calls Palantir a great company and a great platform, but points to cost and vendor lock-in. Cyberhill's counter is malleability and implementation speed: "we can implement it in days, not months."Compete where the playing field is level: Instead of fighting Salesforce for attention in New York, LA, and San Francisco, Cyberhill markets at the local level in secondary markets like Dallas and Denver, with billboards, airport advertising, and quarterly steak dinners. Own the category conversation before the window closes: Rob predicts everybody will be talking about the semantic layer in enterprise AI within three years, if not sooner. He also acknowledges the loudest technology does oftentimes win, which makes owning that conversation early a strategic requirement, not a vanity project. // Sponsors: Front Lines -- Silicon Valley's leading Podcast Production Studio. We help B2B tech companies launch, manage, and grow podcasts that drive demand, awareness, and thought leadership. Mention you are a listener and get a 10% discount. www.FrontLines.io/Podcast-as-a-Service

In a recent episode of BUILDERS, we sat down with ⁠Aaron Fulkerson⁠, CEO of ⁠OPAQUE⁠, to learn how the company found its market in sovereign AI deployments after discovering that customers would accept 90% verifiability with known gaps rather than waiting for perfect end-to-end confidential computation.Topics Discussed:Why Aaron joined OPAQUE from ServiceNow only after the founders agreed to apply the technology to language modelsHow generative AI systems leak data by architecture, and why that threatens both enterprises and frontier labsThe first customer: a European Union cybersecurity agency running analytics and MLHow the focus evolved from high tech to regulated industries to sovereign deployments over roughly sixteen monthsThe 2025 realization that a trust artifact with 90% verifiability and known gaps was good enough for customersTurning down a nation-state-scale sovereign deployment in the UAE, and how that honesty turned the entity into a Series B investorWhy OPAQUE handed the Confidential Computing Summit to the Linux FoundationWhat Apple's private cloud compute expansion signals for enterprise AI adoptionGTM & Technology Adoption Lessons:Refuse to commercialize the wrong product: Before joining, Aaron told the founders that if the product did not involve language models, it was probably not interesting and would be too difficult to build a commercial effort around. The replatforming from confidential Spark analytics to language models started essentially on day one.Let customers define good enough: OPAQUE assumed it needed end-to-end confidential computation across CPUs and GPUs, but confidential GPU availability through hyperscalers was slow. The company discovered through customers that a trust artifact delivering roughly 90% verifiability of policies, with known gaps, was acceptable. Aaron called it a we-have-been-overthinking-this moment.Honesty about scale limits can win the deal: When a UAE entity wanted OPAQUE to roll out as part of a sovereign stack for a 13 gigawatt build-out, Aaron told them nation-state scale was not deliverable in 2025, since the company had just put its first customers into production. OPAQUE took on two or three enterprise-scale projects instead, and the entity became an investor in the Series B round.Follow the pain into regulated industries: The initial focus was high tech, but OPAQUE learned that regulated industries with strict requirements had the real urgency. Just over a year before the conversation, the company shifted focus to sovereign deployments: healthcare data, banking data, and high tech that is critical infrastructure.Build the ecosystem, not just the brand: OPAQUE hosts the Confidential Computing Summit, which grew until the Linux Foundation became co-host and took ownership. Aaron's view is that no one entity can own the digital sovereignty conversation; OPAQUE needs Google, Microsoft, and Apple building interoperability for the category to exist.Use the anchor example customers already trust: Aaron points to Apple architecting Siri's Gemini-powered processing to be confidential end-to-end. If a basic chatbot was too great a data leakage risk for Apple, enterprises running far leakier AI agents on far more valuable data have their answer.// Sponsors: Front Lines -- Silicon Valley's leading Podcast Production Studio. We help B2B tech companies launch, manage, and grow podcasts that drive demand, awareness, and thought leadership. Mention you are a listener and get a 10% discount. www.FrontLines.io/Podcast-as-a-Service

Guild AI is building the infrastructure layer that enterprises need to deploy, manage, and govern AI agents operating inside their systems. Founded by ⁠James Everingham⁠ — a five-time founder who most recently led developer infrastructure at Meta, overseeing a team of more than a thousand engineers — ⁠Guild AI⁠ is solving a problem James watched emerge firsthand inside one of the world's most complex technology organizations: what happens when agents stop being prototypes and start taking autonomous action inside your production infrastructure.In this episode of BUILDERS, James shares what he's learned about founding companies across nearly four decades, from writing shareware in central Pennsylvania in 1987 to building developer tooling at Meta to launching Guild AI. He goes deep on why the go-to-market motion for a new infrastructure category defaults tops-down, what enterprise engineering leaders actually respond to from vendors, and the product philosophy he's refined across five companies.Topics Discussed:Why James founded Guild AI after watching agent adoption break at scale inside MetaHow Guild AI's go-to-market shifted from developer-led to CIO/CSO/CTO-driven — and what forced that shiftWhy design partners pushed Guild AI's use cases away from software development and into legal, marketing, HR, and financeWhat vendors consistently got wrong when trying to sell to James at Meta — and what actually workedGTM Lessons For B2B Founders:New infrastructure categories require a tops-down entry: Guild AI initially assumed a bottoms-up, developer-led motion. Their design partners revealed quickly that CIOs, CSOs, and CTOs were the ones carrying the urgency around agent governance — not individual developers. James frames the dynamic precisely: in early markets where buyers lack a clear starting point, individual contributors won't self-organize around a new tool without leadership buy-in first. The tops-down entry earns executive trust, establishes the governance framework, and creates the conditions for developers to adopt the tooling on their own terms — without a mandate.Follow your design partners, not your original thesis: Guild AI launched with a developer productivity thesis. Their earliest design partners pulled them toward marketing, legal, HR, and finance — areas where, as James noted, there's likely more demand for agentic workflow automation than in software development itself. The team followed the signal rather than defending the thesis. For B2B founders in emerging categories, design partnerships aren't just validation — they're navigation. The customers who show up first will often redirect you toward the higher-value problem you hadn't fully seen yet.Enterprise outreach is pattern-matched and rejected in seconds: James was receiving vendor pitches at Meta at scale. What killed deals before they started: AI-generated emails that reflected his own company's marketing language back at him with light personalization layered on top. His analogy is precise — he compares it to the first banner ads on the internet, effective for about three months before everyone learned to ignore them. // Sponsors: Front Lines — We help B2B tech companies launch, manage, and grow podcasts that drive demand, awareness, and thought leadership.⁠ www.FrontLines.io⁠The Global Talent Co. — We help tech startups find, vet, hire, pay, and retain amazing marketing talent that costs 50-70% less than the US & Europe.⁠ www.GlobalTalent.co⁠//Don't Miss: New Podcast Series — How I Hire Senior GTM leaders share the tactical hiring frameworks they use to build winning revenue teams. Hosted by Andy Mowat, who scaled 4 unicorns from $10M to $100M+ ARR and launched Whispered to help executives find their next role. Subscribe here:⁠ https://open.spotify.com/show/53yCHlPfLSMFimtv0riPyM⁠

Two months before this conversation, The Biological Computing Co (TBC) was still in stealth. Three years of building, rebranding, and accumulating experimental data — all before showing the world a single thing. In this episode of BUILDERS, we sat down with ⁠Alexander Ksendzovsky⁠, Co-Founder and CEO of ⁠TBC⁠, to hear how a neurosurgeon-scientist turned a 20-year research obsession into a commercial AI optimization company. TBC grows living neuron cultures, connects them via electrodes, and derives software tools from the biology's computation — currently applied to video generation model optimization, with language models and world models on the roadmap. The ICP they're selling to today is not the one they envisioned at founding. The name they operate under is a deliberate category ownership move. And the stealth-to-launch decision was made entirely on one criterion: enough data to show the world, not just tell it.Topics Discussed:How TBC uses living neuron cultures and electrode arrays to derive optimization techniques for AI video generation models that silicon-based approaches can't replicate Why TBC's current ICP — video gen model companies that have exhausted standard optimization techniques — is not the customer they anticipated when they founded the company TBC's two-track marketing framework: awareness and credibility, and why sequencing them correctly matters for deep-tech GTM The stealth-to-launch decision: the specific data threshold that triggered TBC's emergence after three years, and why they would not have come out earlier How investor pitch feedback — not customer feedback — drove TBC's rebrand from academic positioning to product-forward identity Category creation in a nascent field: TBC's approach to building the ecosystem rather than fighting for definitional control How TBC structures hiring and team cadence to keep exploratory research culture from blocking commercialization GTM Lessons For B2B Founders:Time your launch to proof, not momentum. TBC spent three years in stealth before going public — not because they feared competitors, but because they needed enough experimental data to generate belief rather than just curiosity. Alex's framing: "If we had done this two years ago, we had some interesting experiments, but people would think it's just research at that point. We really needed people to see that we're building real tools, they're productized and they're ready for production now." For deep-tech founders, the stealth exit decision isn't about market timing — it's about whether your evidence base crosses the threshold from interesting to credible.Investor pitch feedback is your earliest positioning stress test. TBC's rebrand wasn't triggered by customer research — it was triggered by investors consistently not understanding what they were building. Both founders came from academic neuroscience and were unconsciously pitching in that register. Alex: "We learned very early on through pitching to investors that the way we were positioning it was way too academic. We had to beat that out of ourselves." For pre-revenue founders, if investors with context can't quickly grasp your value proposition, buyers with less context won't either. Fix the positioning before you scale the outreach.// Sponsors: Front Lines — We help B2B tech companies launch, manage, and grow podcasts that drive demand, awareness, and thought leadership.⁠ www.FrontLines.io⁠The Global Talent Co. — We help tech startups find, vet, hire, pay, and retain amazing marketing talent that costs 50-70% less than the US & Europe.⁠ www.GlobalTalent.co⁠//Don't Miss: New Podcast Series — How I Hire Senior GTM leaders share the tactical hiring frameworks they use to build winning revenue teams. Hosted by Andy Mowat, who scaled 4 unicorns from $10M to $100M+ ARR and launched Whispered to help executives find their next role. Subscribe here:⁠ https://open.spotify.com/show/53yCHlPfLSMFimtv0riPyM⁠

Most AI companies in 2023 raced to own a vertical. EverWorker made the opposite bet — build a horizontal platform that lets anyone create agents for any purpose, no code required. In this episode of BUILDERS, ⁠Anton Antich⁠, CPO and Co-Founder of ⁠EverWorker⁠, gets into what it actually takes to sell AI inside enterprises that are stuck between the hype and the reality, why he's making the case against SaaS entirely, and how an early PLG motion gave way to deep consultative selling once they realized the market wasn't where Silicon Valley thought it was. Anton helped scale a company from $0 to $1B ARR — and he's direct that most of what he learned there no longer applies.Topics Discussed:Why the $0–$1B scaling playbook is obsolete in the AI eraEverWorker's pivot from PLG to enterprise consultative sellingTargeting "boring billion-dollar companies" as a deliberate ICPWhy most AI pilots never reach production — and the services motion that fixes itThe org-chart model for AI agent teams and the "Chief of Staff" productThe case for replacing SaaS entirely with agents, databases, and markdown filesGoing down-market in 2026 and why community is the lead growth channelWhy instant product access has replaced "contact us for a demo" as the conversion standardGTM Lessons For B2B Founders:Audit your team's DNA before choosing your GTM motion. EverWorker launched PLG, then quickly realized their entire founding team came from enterprise — Microsoft, VMware, Veeam. The pivot wasn't a failure; it was an honest read of where their unfair advantages actually lived. Before committing to a motion, map your team's network, sales instincts, and domain depth. Those signals will outperform market trend-chasing every time.Build a services layer or watch your pilots die. The gap between AI pilot and production is where most deals go to die — Anton cites the widely-reported stat that the vast majority never make it through. EverWorker's solution was to build a services organization that identifies two or three mundane, high-friction processes — Anton's example is data entry, work humans find demeaning and AI handles well — automates them fast, and uses that visible win to build organizational trust. The services layer isn't a concession. For complex AI sales right now, it's the mechanism that actually converts pilots into production.Your ICP should be defined by who won't default to "we'll build it ourselves." EverWorker learned this the hard way in enterprise. Walk into a Fortune 500 or a tech-forward company and IT shows up in the room and kills the conversation. Anton's team shifted toward what he calls "boring billion-dollar companies" — industries doing real, essential work that don't get the spotlight and can't afford to staff AI expertise internally. These buyers need the outcome, not the platform, and they don't have an internal team to rationalize building around. That dynamic is a structural GTM advantage.// Sponsors: Front Lines — We help B2B tech companies launch, manage, and grow podcasts that drive demand, awareness, and thought leadership.⁠ www.FrontLines.io⁠The Global Talent Co. — We help tech startups find, vet, hire, pay, and retain amazing marketing talent that costs 50-70% less than the US & Europe.⁠ www.GlobalTalent.co⁠//Don't Miss: New Podcast Series — How I Hire Senior GTM leaders share the tactical hiring frameworks they use to build winning revenue teams. Hosted by Andy Mowat, who scaled 4 unicorns from $10M to $100M+ ARR and launched Whispered to help executives find their next role. Subscribe here:⁠ https://open.spotify.com/show/53yCHlPfLSMFimtv0riPyM⁠