The Next Unicorns: 7 Top AI startups from the HF0 Residency

A car is relatively easy to imagine: take the wheels of a horse-drawn vehicle, replace the horse with an engine, and you have motorized transport. Traffic is harder to foresee. Diego Rodriguez, Krea’s co-founder and CTO, uses that distinction to ask what follows from cheap AI creation. The immediate achievement is generating images; the downstream problem is finding anything useful among them. Rodriguez reports that Krea generates a million images per day for one studio. Creative abundance creates a discovery problem.

His second story is the Tower of Babel: a shared ambition collapses when people can no longer understand one another. The modern equivalent is a stand-up meeting where one person insists on React and another on JavaScript, without resolving what they are trying to communicate. For Krea, an AI creative suite, the purpose of the tools is to help people convey ideas and tell stories.

The third story returns to discovery. In a conversation with someone at Netflix, Rodriguez hears the question of how viewers will find content when it is personalized to individual towns in India. Then the speed of production becomes concrete: Rodriguez reports that a customer went from signing up for Krea through conversion, payment and a Fox advertisement broadcast to millions of people in two days. Creation, distribution and discovery are becoming tightly connected product problems.

Rodriguez closes by reporting 25 million users with an eight-person team. He mentions funding without an amount, shows customer examples, and offers a dedicated email address to prioritize applications from the event. The recruiting invitation follows directly from the scale problem: a small team is building the tools around a rapidly expanding volume of creative work.

OpenHome begins with a quiz: after smartphones and laptops, what was the previous year’s third best-selling consumer product? An audience member guesses Apple Watch; the presenter answers smart speakers. He claims 500 million smart speakers were sold in the prior year, although the pitch supplies no market dataset establishing annual sales. The product complaint is easier to recognize: these devices are difficult to talk to, offer little customization, and lack a community that can extend them.

OpenHome presents itself as the first AI-driven smart speaker, with a broader goal than replacing one household device. Users should be able to chat naturally instead of learning command language, while developers should be able to put that interaction into whatever physical form they choose.

The presenter’s case for an ecosystem comes from earlier work: chief of staff to Splunk’s founder, a role on MakerDAO’s founding team, and a data-privacy business that he and his co-founder later sold. He describes Splunk as a $30 billion company, MakerDAO as a $5 billion developer ecosystem, and the prior business as having raised $50 million. The connecting lesson is that developers determine which applications people actually want.

The presenter reports more than 10,000 developers building on OpenHome. Its developer proposition combines three elements:

• Open source: developers can participate in extending the platform.
• LLM-driven interaction: conversation becomes the interface to applications.
• Device freedom: the pitch’s “fully jailbroken” positioning emphasizes customization, including talking toys, AI robots and appliances.

The dashboard and editor bring that proposition into a development environment. The presenter describes hundreds of applications, including games, personalities and home-automation tools.

The hardware offer makes the invitation tangible. After saying the previous dev-kit allocation was booked within minutes, the presenter announces a new batch of 500 free dev kits for attendees, with shipping included in the offer. Developers are being invited to build their own smart speakers and voice applications on the devices, not merely try a finished assistant.

Josh, the founder of Coframe, arrives at website optimization through autonomous coding. He reports that his previous company grew to more than $2 billion in a couple of years, and that an autonomous coding agent he created ranked number one on GitHub for a week. That experience led him toward a larger application of code generation: websites that can change their customer experience instead of remaining static.

A conventional website serves a largely fixed, one-size-fits-all experience. Coframe’s proposition is to give each customer experience its own AI growth team, making the site adaptive and personal. The pitch stays at the product level rather than demonstrating how variants are generated or experiments are assigned, but the intended shift is clear: the website becomes something that can continually improve its presentation.

Josh reports that Coframe generated $20 million for Europe’s largest travel company in a few weeks. He also reports a click-through-rate increase within a few weeks for India’s largest company, which he describes as a $400 billion enterprise; no improvement magnitude is given. The pitch does not supply the attribution method or experimental conditions for those customer results. He closes by claiming that Coframe was then the only marketing-tech company directly partnered with OpenAI, recounting OpenAI’s praise for the team, and inviting interested attendees to reach out.

Eugene opens with an architecture claim. His team built Qwerky-72B, which he describes as the largest model without transformer attention, using eight GPUs. He claims a thousandfold reduction in inference with equivalent performance, but does not identify whether the measured quantity is latency, cost, memory or another metric, nor give the baseline and workload. The associated model work concerns conversion of a pretrained model, so the eight-GPU statement should not be read as training a 72-billion-parameter model from scratch. Eugene then says the technology can also help existing transformers through speculative decoding.

His objection to scaling is that more intelligence does not automatically make an agent dependable. He contrasts spending billions on larger models with unresolved errors in chains of actions. Eugene attributes a more-than-ten-year horizon for fixing compound agent errors to DeepMind’s founder and CEO, invokes Yann LeCun’s call for a new architecture, and claims that more than 90% of production AI projects fail. These are motivations offered in the pitch, without their underlying sources or study definitions. The engineering problem he identifies is reliability.

Consider his hypothetical delivery app that succeeds only 45% of the time. Sometimes the order disappears; sometimes it produces 100 pizzas. The user is left dealing with customer support and cleaning up the mistake. An impressive successful run does not compensate for that experience. Eugene applies the same test to agents built on frontier models: can they reliably book an airline ticket, sort email, or handle taxes and invoices? Those jobs require dependable execution more than PhD-level mathematics.

Featherless AI describes its research goal as personalized AGI made reliable for individual users. Eugene says Action-R1 outperforms Claude 4 Sonnet, Gemini and OpenAI models on the agent and form-filling work he discusses. The pitch supplies neither the evaluation protocol nor the precise Gemini and OpenAI model versions, so its language of absolute reliability does not establish universal task completion. His subsequent 99.9% reliability figure is a goal for routine tasks, not a demonstrated result.

Reliability is revenue is the commercial conclusion. Each everyday task that becomes dependable opens a possible application in areas such as e-commerce or B2B sales. Eugene frames these as potential billion-dollar businesses built around useful actions, rather than another increment in abstract model intelligence, and invites interested attendees to contact him.

Jonas Bauer introduces Upside through a long-standing interest in data. He left high school, moved to California and joined Branch, whose mobile links connected users to apps. Jonas reports leading a Branch team that built a search engine used by more than 100 million people each day. He left the previous year with one of Branch’s founders to work on forensic revenue attribution and intelligence.

The problem appears in an ordinary inbox: a sales email that its recipient has no intention of answering. Sales and marketing teams keep sending messages because they do not know which activities work. They have plenty of data, but storing it is not the same as understanding it. Jonas describes Salesforce as “a SQL database in a trench coat”: a place where teams accumulate records without necessarily knowing how to use them.

LLMs can turn the contents of a record into explicit fields. Jonas’s example is extracting the important details from a poorly handled email record into structured form. A compact illustrative JSON record shows the distinction between retaining the original message and adding an interpretation:

{
  "record_id": "email-001",
  "body": "Hi Maya, would you like a demo of our analytics tool?",
  "extracted": {
    "recipient_name": "Maya",
    "interaction_type": "sales_outreach",
    "offering": "analytics tool",
    "requested_action": "schedule_demo"
  }
}

The example’s message and field names are illustrative. The useful operation is extraction: the text remains available while its important details become addressable data. The message alone does not establish that Maya replied, booked a demo or became a customer.

The next step is to connect those structured records. Just as search engines and crawlers made the unstructured web navigable, Upside aims to turn raw enterprise data into a map of people and their interactions. That map is the foundation for a data command center through which teams can understand their activity and reach customers more effectively. The progression is from stored text, to structured interactions, to a usable representation of the business.

After roughly a year of quiet development, the team began discussing the product publicly a couple of weeks before the pitch. Jonas says a co-founder’s LinkedIn post generated a wave of demo requests. The displayed engagement slide accompanies that account of demand; it is evidence of attention, while the customer queue he describes is prospective access to the platform. With more building ahead, he recruits for knowledge graphs, data analytics agents, graph analytics and graph learning models.

Shijia introduces himself as the founder of OpenAI, then immediately corrects the joke to OpenAudio. He previously created Fish Audio. Shijia reports growth from $400,000 to $5.5 million in annualized revenue in four months, and a seed round at a $100 million valuation.

His origin story is deliberately personal: a six-year relationship stretching from high school to college ended after he discovered infidelity. He recounts asking how he could trust relationships again and arriving, to audience laughter, at AI. But the product obstacle is the voice. Flat, emotionless, robotic speech makes an AI companion difficult to connect with. He therefore started with speech, naming so-vits-svc, Bert-VITS2 and Fish-Speech as the team’s open-source work.

Shijia introduces OpenAudio S1 as having launched two days before the presentation, correcting his initial statement that it launched that day. He positions it as the first instructable voice model: the user controls both what is said and how it is delivered. In the demo, visible parenthetical cues include emphasize and whispering; the spoken sequence also moves into a betrayal-themed exclamation. Delivery becomes part of the input, alongside the words to be spoken.

The demo returns to the relationship story with the joke that the voice will never leave. Shijia then claims S1 leads ElevenLabs on TTS Arena and characterizes the competitor’s release that day as only a demo. No dated leaderboard snapshot or evaluation conditions accompany that comparison. He closes with the practical distinction he wants listeners to remember: Fish Audio is available to try immediately.

David Vorick introduces Glow after recounting his work on Siacoin. Vorick reports that Siacoin grew from a $10,000 market capitalization to more than $3 billion, and that Framework and Union Square Ventures led a $30 million round into Glow. He then claims an on-chain DePIN revenue record of more than $10 million in a single day. Those figures establish the scale of his pitch, although the revenue definition and record comparison are not supplied.

Glow’s mechanism is to build solar through incentives. Vorick points to a photograph that he says his team took in India: a solar farm constructed for the purpose of mining Glow tokens. Token rewards are intended to make building solar capacity economically attractive, redirecting the motivation behind mining toward electricity generation.

The need is physical. Vorick describes rising temperatures and growing populations straining grids in the developing world. When people cannot run air conditioning during the hottest part of the day, heat becomes a threat to life. He claims Glow’s incentive protocol can turn the same government subsidy into ten times as much solar, without supplying the subsidy baseline or calculation.

Vorick names incentive projects in India, Mexico and Lebanon, with broader global ambitions. His closing comparison is Bitcoin: if token incentives could cause people to construct tens of millions of mining machines, why not use incentives to produce tens of millions of solar panels? He invites interested collaborators to get in touch.

The next presenter also introduces himself as David, this time for Favorited. He reports that an earlier social app reached 250 million users and made $20 million a year. His setup is that others attributed the success to luck, so he set out to repeat it.

David describes Favorited as the world’s most engaging live app and reports growth from $1 million to $100 million annualized in six months. The short pitch supplies no engagement definition or detail about that annualized measure. It closes as a direct recruiting appeal to strong engineers, with David calling the business the fastest-growing company of all time and inviting attendees to talk to him.

Alex Atallah introduces OpenRouter as the first and largest LLM marketplace. Having co-founded OpenSea in 2017, he became interested in a different market at the end of 2022: would inference be winner-take-all? The first experiment was a Chrome extension that let people bring their own language model to websites supporting its protocol. That experiment evolved into a single API for accessing language models.

The common interface is meant to remove work that would otherwise recur for each integration. A developer has one payment relationship and can change models with what Atallah describes as near-zero switching costs. OpenRouter handles tool-calling differences, edge cases and caching, while seeking favorable prices, performance and uptime for the user’s region or server location.

Model discovery also needs to reflect the requirements of an application. Atallah names marketplace filters for context, features, tool calling and structured output:

These filters make the marketplace useful for selecting a compatible model, rather than simply browsing model names.

OpenRouter then adds a chat room for comparing models head-to-head, with an interaction Atallah likens to iMessage. Fine-grained privacy settings include API-level controls. Observability shows which models are being used and why, while public rankings expose real-world usage and categories of prompts. Those rankings describe use of the marketplace, rather than a single benchmark of model quality.

Atallah reports monthly growth of 10–100% over two years, correcting an initial reference to two months; he does not identify the metric being measured. Scaling the system has consequently become a substantial part of the work. The underlying goal is to make a heterogeneous provider ecosystem feel homogeneous. In his example, Claude through Bedrock should behave like Claude through Vertex or directly through Anthropic, with OpenRouter handling the abstraction between them.

The final technical detail is OpenRouter’s custom inference middleware, called plugins. Atallah compares plugins with MCP integrations, then identifies two capabilities: a plugin can call MCPs internally, and it can transform a language model’s output. That places the middleware around inference itself, where it can connect to other services and change the returned result.

Atallah also claims the fastest routing in the market, without presenting a latency benchmark. He lists images, enterprise features and prompt observability as work planned for the coming months; these are roadmap items at this point in the recording. He ends by inviting attendees to meet him afterward or visit the careers page. The engineering invitation is specific: normalize access to diverse providers, route requests efficiently, and build the middleware and visibility that applications need around model inference.