Karumi: Revenue, team and how it was built
Karumi (AI agents that give live product demos of business software over a video call, USA (New York), co-founder from Spain): Reported US$31K MRR three weeks after launch, March 2026. Evidence grade D: reported, not verified. Team: 2 co-founders, Pablo Omenaca (CEO) and Toni López (CTO); 4 people in March 2026 (Valencia Plaza) and 6 on the Y Combinator profile in October 2026; headcount at the headline figure not published. Launched November 2025; still running.Added 27 October 2026
- Outcome
- Reported US$31K MRR three weeks after launch, March 2026
- Evidence
- D reported, not verified
- Launch
- November 2025
- Status
- Still running
- Team
- 2 co-founders, Pablo Omenaca (CEO) and Toni López (CTO); 4 people in March 2026 (Valencia Plaza) and 6 on the Y Combinator profile in October 2026; headcount at the headline figure not published
- Outside funding
- Y Combinator (Fall 2025); a directory estimates US$2M raised in total; co-founder Toni López said in March 2026 that they plan a Series A of 15 to 20 million
- Main channel
- Not stated
What is not verified
The headline figure, US$31,000 in monthly recurring revenue three weeks after launch, rests on one statement by co-founder Toni López, reported by Estefanía Pastor in Valencia Plaza on 29 March 2026. The article does not say on which date the figure was reached, only that it was during the Y Combinator program of Fall 2025. No payment provider, investor or second interview confirms it, and no later revenue figure was found. A reader could look for a funding announcement that states revenue, a Y Combinator Demo Day report, or a payment-provider profile. The headline number is grade D: one claim or one estimate, with no second source. Aiolum shows it because the story is useful, and marks it so that it is not mistaken for a verified figure.
Karumi makes AI demo agents: programs that join a video call with a person who wants to see a software product, share a screen, use the real product in a web browser, and explain it by voice while the person asks questions. A product demo is the walkthrough that a salesperson usually gives before a company buys business software. The agent combines a large language model (LLM), a model trained on large amounts of text that can write and answer questions, with speech-to-text and text-to-speech models and a browser agent, a program that clicks, types and reads a web page as a person would. Karumi's customers put it on their own website in place of a request-a-demo form, or use it for onboarding and support.
Toni López and Pablo Omenaca met at StackAI, a company of the Winter 2023 batch of Y Combinator (YC), a start-up accelerator that invests in young companies and runs a three-month program for them. They started Karumi in 2025 and joined YC's Fall 2025 batch. López told the Valencia newspaper Valencia Plaza, in an article published on 29 March 2026, that the first idea was a tool to automate the communication of product updates, that the founders changed to demos during the program, and that after the launch they reached US$31,000 in monthly recurring revenue (MRR, the monthly value of all active subscriptions) in three weeks, with about 10,000 demos. The same article says that the team was four people and had moved from San Francisco to New York.
The headline figure is grade D: it is one founder statement, reported once, with no date for when it was reached and no later revenue figure in any source read. Other numbers are easier to check: in October 2026 the home page counted 102,023 demos delivered and the Y Combinator profile listed six people. López said in March 2026 that the founders want to raise a Series A round of 15 to 20 million in the United States. The page also shows five failed companies that tried to scale sales conversations with hired callers, live chat, pop-ups, lead scores or chatbots.
How much revenue does Karumi make?
Each row shows the evidence grade and the source numbers from the list at the end of this page. The grades are explained in How Aiolum verifies.
| Fact | Value | Grade | How we know | Sources |
|---|---|---|---|---|
| Monthly recurring revenue three weeks after launch, stated in March 2026 | US$31,000 | D | Valencia Plaza, 29 March 2026, reporting what co-founder Toni López said: "Según explica, alcanzaron 31.000 dólares de ingresos recurrentes mensuales (MRR) en tres semanas" [As he explains, they reached 31,000 dollars in monthly recurring revenue (MRR) in three weeks] (translation). The figure is in US dollars in the source. One source and one statement; no payment-provider or investor confirmation. | [1] |
| Yearly run rate from that MRR | about US$372,000 | calc | Calculated: 31,000 times 12. This is the outcome amount stored for this page. No source states a yearly figure. | [1] |
| When the figure was reached | late 2025, during the Fall 2025 batch of Y Combinator | calc | Calculated, not stated. The article says that Karumi launched its product during the program and reached the figure in three weeks. The Launch YC post is dated 10 November 2025 in the page data; 10 November 2025 plus 21 days is 1 December 2025. | [1, 2] |
| Time from launch to the headline figure | three weeks; stored as 1 month | calc | Calculated: 21 days is about 0.7 months, rounded up to 1. | [1] |
| Demos in the launch period | about 10,000; 2,000 on the first day | C | Valencia Plaza: "realizaron alrededor de 10.000 demostraciones. El primer día llegaron a registrar 2.000 demos" [they gave about 10,000 demos. On the first day they recorded 2,000 demos] (translation). The article does not say on which date the 10,000 count was taken. | [1] |
| Demos delivered, 27 October 2026 | 102,023 | C | Counter on the home page: "102,023 demos delivered". A company figure, not checked by anyone else. | [3] |
| Demos in October 2026 compared with the launch-period count | about 10 times | calc | Calculated: 102,023 / 10,000 = 10.2. This counts demos, not paying customers or revenue. | [1, 3] |
| Founders | Pablo Omenaca (CEO) and Toni López (CTO) | C | Y Combinator profile. Both worked at StackAI (YC W23) before: Omenaca as COO and founding engineer, López as founding engineer; López also worked as a software developer at IBM Research. Valencia Plaza says López was born in Meliana. | [2, 1] |
| Y Combinator batch and founding year | Fall 2025 (F25); founded 2025 | C | Y Combinator profile: "Batch: Fall 2025", "Founded: 2025", status Active. | [2] |
| Team, March 2026 | 4 people | C | Valencia Plaza: "el equipo de Karumi está formado por cuatro personas" [Karumi's team is made up of four people] (translation). This is the team size closest to the headline figure in any source read. | [1] |
| Team, October 2026 | 6 on the Y Combinator profile; 7 open roles | C | Y Combinator profile: "Team Size: 6". The about page lists seven open roles in New York, Madrid and San Francisco. A startup directory gives 10 employees in a profile last changed on 2 July 2026 (grade D). | [2, 4, 10] |
| Outside funding | Y Combinator; about US$2 million in total by a directory estimate | D | The StartupHub.ai profile says "Total funding: $2M". No source read names the investors or the amount of any round other than the Y Combinator investment, whose amount no source read states. | [10, 2] |
| Planned Series A | 15 to 20 million, currency not stated, not yet raised in any source read | C | Valencia Plaza: "Nos gustaría levantar entre 15 y 20 millones en Estados Unidos" [We would like to raise between 15 and 20 million in the United States] (translation). | [1] |
| Revenue goal stated in March 2026 | 1.5 to 2 million a year, currency not stated | C | Valencia Plaza: "alcanzar entre 1,5 y 2 millones de facturación anual" [reach between 1.5 and 2 million in annual billing] (translation). A goal, not a result. | [1] |
| Goal compared with the late-2025 run rate | 4.0 to 5.4 times | calc | Calculated: 1,500,000 / 372,000 = 4.0 and 2,000,000 / 372,000 = 5.4, if both figures are in US dollars. | [1] |
| Customers | large companies, including one in the S&P 500 index | C | Valencia Plaza: "Karumi ya ha cerrado acuerdos con grandes compañías, incluida una empresa del índice S&P 500" [Karumi has already closed deals with large companies, including one company in the S&P 500 index] (translation). The company is not named. Karumi's own customer story names Salesforge. | [1, 8] |
| One customer's result | about 300 demos a month (+46%) and 21% conversion to a trial | C | Karumi's Salesforge customer story: "~300 demos/month (+46%)", "21% conversion to trial". Written by Karumi, with a quote from the customer's CEO. | [8] |
| How the agent is presented to prospects | as an AI demo agent | C | Home page: "Talk to our AI demo agent." Launch YC post: "a live, interactive video call with an AI agent". The Terms of Service describe "AI agents that join video calls to deliver product demonstrations". | [3, 2, 9] |
| Time to set up for a customer | less than a week | C | López in Valencia Plaza: "Se tarda menos de una semana en implantarlo" [It takes less than a week to set it up] (translation). | [1] |
| Prices | not published | C | Pricing page: two plans, Growth and Enterprise, both "Talk to us". | [5] |
How to read this table: a grade C value is a statement by the founders or the company, on its own pages, on its Y Combinator profile or to a journalist. A grade D value rests on one source that no second source confirms, or on a directory estimate. An "X" value was computed on this page from the rows above it. Quotations in Spanish are given in the original, followed by this page's English translation.
| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| Growth | custom | custom | Custom agents, languages and sessions; agent builder; session insights; data export to the CRM; MCP connection |
| Enterprise | custom | custom | Everything in Growth, plus on-premise deployment, SOC 2 Type II, ISO 27001 and GDPR, service-level agreements, dedicated server capacity and a dedicated success team |
The pricing FAQ says: "Pricing depends on the number of agents, expected demo volume, languages and integrations, so we scope it with you on a short call." Most teams start with a pilot on one use case. No source read for this page states a price, so the rows show the plan contents only.
When did Karumi launch, and how did it grow?
From two engineers who left StackAI in 2025 to a demo agent that, by the co-founder's account, reached US$31,000 a month three weeks after launch. Each entry carries its grade.
- Before 2025CToni López and Pablo Omenaca work at StackAI, a Y Combinator company of Winter 2023: López as a founding engineer, Omenaca as COO and founding engineer. Valencia Plaza says López had earlier built an AI avatar app for iPhone and founded Cactus AI in San Francisco. [2, 1]
- 2025CThey found Karumi and join Y Combinator's Fall 2025 batch. The first idea is a tool that automates the communication of product updates. [2, 1]
- During the YC program, 2025CThey drop the first idea as too complex to integrate for an early company and build an agent that only has to give a demo. [1]
- 10 November 2025CLaunch YC post: "Karumi - Your AI demo agent. 24/7." It invites readers to try a live demo and names Deel's platform as one place where the agent runs. [2]
- About three weeks after the launch, late 2025DLópez later says the product gave 2,000 demos on its first day and reached US$31,000 in monthly recurring revenue in three weeks, with about 10,000 demos. [1]
- 28 December 2025CLast update of the Terms of Service, which describe AI agents that join video calls to give product demonstrations. [9]
- 29 March 2026CValencia Plaza publishes the founder's account: a team of four, a move from San Francisco to New York, a goal of 1.5 to 2 million in annual billing, and a planned Series A of 15 to 20 million. [1]
- 2 July 2026DA startup directory profile, last changed on this date, gives 10 employees and US$2 million of funding in total. These are estimates. [10]
- 27 October 2026CThe home page counts 102,023 demos delivered and adds a Product Brain that gives other AI agents the customer's product knowledge through an API or MCP (Model Context Protocol, a standard way for AI assistants to call outside tools). The Y Combinator profile lists six people; seven roles are open. [3, 2, 4]
Why did Karumi work?
Four causes, in the order they mattered. Each one is a reusable lesson.
A demo does not need the whole product
López told Valencia Plaza: "Vimos que entender todo un producto era muy complejo, pero para hacer una demo no necesitas conocerlo en profundidad" [We saw that understanding a whole product was very complex, but to give a demo you do not need to know it in depth] (translation). The founders kept the hard part, an AI that works inside real software, and cut the scope to one job with a clear buyer.
A video call, not a recording
The prospect clicks a link and, by the home page's account, the agent shares its screen within two seconds. It opens tabs, navigates the customer's product and answers spoken or typed questions, at any hour and in several languages. A recorded video cannot answer a question, and a salesperson is not available at every hour.
Sold where demo requests pile up
The launch post addresses software companies whose inbound demo requests grow faster than their sales staff. López said that a customer "puedes eliminar el 90-95% de las llamadas de demos" [you can remove 90 to 95% of demo calls] (translation). That is the founder's claim; the one customer story read reports more demos and a 21% conversion to a trial, not a cut in calls.
Data from every demo
Every session produces a transcript and a list of questions, objections and signs of interest that go into the customer's CRM. Valencia Plaza notes that this kind of data is hard to get from pre-recorded videos.
Y Combinator as the first channel
The product launched to the Y Combinator network, and López said it gave 2,000 demos on its first day. The article says the program also gave the founders access to investors and contacts. No source says where most paying customers came from, so this page records the channel as unknown.
How it was built
These details come from Karumi's job postings on Y Combinator's site, its customer story, its pricing page and the Valencia Plaza article. They are grade C, except the hosting, which is visible in the page source (grade B). [6, 7, 8, 1, 5, 3, 2]
- A web app with a React front end and a back end in asynchronous Python with FastAPI, PostgreSQL and Redis, deployed on AWS, according to the Product Engineer posting.
- The AI engineering posting describes the core: speech-to-text and text-to-speech models for the voice, browser agents that navigate web applications using computer vision and the page's DOM (the Document Object Model, the structure of elements that a browser builds from a web page), prompt systems and LLM workflows, and real-time pipelines in which delay matters.
- Training material for each customer: the website, the documents and recordings of real demos. The agent logs into a real demo environment of the customer's product, the Salesforge story says. López said that setup takes less than a week.
- Transcripts, summaries, qualification data and custom fields sync to CRMs such as Salesforce and HubSpot, according to the pricing FAQ.
- The marketing site is served through Vercel, which the page source shows.
- Team: two founders at launch, four people in March 2026, six on the Y Combinator profile in October 2026.
Tools named in the sources
Each one opens its page in the Tools tab, with every story that uses it.
Who tried this before Karumi, and why did they fail?
Five companies in the Loot Drop database tried to scale sales conversations with hired callers, live chat, pop-ups, lead scores or chatbots. The values below are from that database and are not independently verified.
VOIQ
The promiseOutsourced sales calls made by a network of trained human agents, so a company did not need its own sales team.
Why it diedEvery call still needed a paid person. Automated sales tools and freelance marketplaces offered the same capacity for less.
RightHello
The promiseLive chat for online shops that turns visitors into buyers with well-timed conversations.
Why it diedA person had to answer each chat, and Intercom, with far more funding, won the market for chat tools.
Ve Interactive
The promiseSoftware that catches shoppers who are about to leave a site and brings them back with personalized pop-ups and emails.
Why it diedIt was paid only when a sale happened and spent heavily on growth, so the costs grew faster than the income.
Toplyne
The promiseAI that predicts which free users of a software product will pay, so salespeople call the right ones.
Why it diedSales teams shrank in the layoffs of 2022 and 2023, and the tool could not prove a clear return on its price.
Vanti
The promiseAI chat that guides shoppers through a personalized buying conversation.
Why it diedThe conversational AI market was crowded, each customer needed complex integration work, and the cost of winning customers stayed high.
The failed startups come from Loot Drop, a free database of dead startups, retold here in our own words. Loot Drop describes its entries as AI-assisted summaries of public sources, so these cards carry no evidence grade.
What changed
VOIQ and RightHello scaled sales conversations with people, so every extra conversation cost an extra hour of someone's work. Ve Interactive and Toplyne used software to decide when or with whom to talk, but the conversation itself was a pop-up, an email or a call by a salesperson. Vanti tried to automate the conversation with the chat technology of its time and, by the Loot Drop record, failed on integration work and the cost of winning customers.
Karumi's agent talks, listens and uses the real product in a browser at the same time; its job postings describe speech models, browser agents and LLM workflows joined in real-time pipelines. López said the agent gave 2,000 demos on its first day, so on this page's reading the extra cost of one more demo is computing time, not a salesperson's hour. The integration problem has not gone away: the founders dropped their first idea because it needed too much integration, and each customer still needs a demo environment and a setup that López said takes less than a week.
Can one person copy Karumi now?
The narrowing, yes. The start inside Y Combinator, no.
The method that can be copied is the one the founders described: take an AI capability that is too hard to sell in general, here understanding a whole software product, and cut it down to one job with a buyer who already pays people to do it. The product itself is hard to copy alone: it needs real-time voice, browser control and a setup for each customer, and the founders came from a funded start-up and launched to the Y Combinator network. Competitors already sell interactive and AI demos; Karumi's own comparison pages name 1mind, Storylane, Navattic, Supademo, Arcade and Consensus.
Where it is still open: Narrower versions are still open: a demo agent for one industry, one type of software or one language market, sold to companies that are too small to pay for an enterprise contract.
- 1Find a sales or support task that companies pay people to repeat many times a day, and that follows the same pattern each time.
- 2Cut the AI's job to that task, and drop the parts that need deep integration with each customer.
- 3Show the real product live instead of a recording, and let the prospect interrupt and ask.
- 4Send what the AI learns in each session to the tools the customer already uses, such as the CRM.
What could go wrong for Karumi?
The risks are the mirror image of the causes above.
What is not verified
The headline figure, US$31,000 in monthly recurring revenue three weeks after launch, rests on one statement by co-founder Toni López, reported by Estefanía Pastor in Valencia Plaza on 29 March 2026. The article does not say on which date the figure was reached, only that it was during the Y Combinator program of Fall 2025. No payment provider, investor or second interview confirms it, and no later revenue figure was found. A reader could look for a funding announcement that states revenue, a Y Combinator Demo Day report, or a payment-provider profile.
An AI that is meant to feel human
López ended the interview with "Nuestro objetivo es que la IA se sienta cada vez más humana" [Our goal is that the AI feels more and more human] (translation), and the launch post says the agent "Matches human reps in clarity and engagement". Karumi's own pages present it as an AI demo agent ("Talk to our AI demo agent"). How each customer labels the agent on its own website was not checked.
Venture funding and a growing team
López said in March 2026 that the founders want to raise 15 to 20 million in the United States. No source read reports that the round has closed. The team grew from four people in March 2026 to six on the Y Combinator profile in October 2026, with seven open roles, and a directory estimate gives ten employees. A large round would turn Karumi into a venture-funded company with a sales team rather than a small team.
Claims on the site that this page could not check
The home page quotes a director at Workday, shows a 4.9 out of 5 rating on G2 and lists results such as "3.2x More meetings vs static assets". These are the company's own statements, not checked by this page.
Integration and competition
The founders dropped their first idea because it needed too much integration work. Each demo still needs a working demo environment of the customer's product, and the comparison pages name six competitors that sell interactive or AI demos.
Sources
The grade next to each source is the best grade that the source can support. "Read in full" means that the text of the page was downloaded and the numbers were checked in it.
- C Valencia Plaza, 29 March 2026, by Estefanía Pastor: Karumi, la startup con gen valenciano que hace demos con IA y ha conquistado a Y CombinatorRead in full in Spanish, in the copy downloaded on 27 October 2026; quotations are translated on this page. Source for the MRR statement, the demo counts, the pivot, the team of four, the revenue goal and the planned Series A.
- C Y Combinator: Karumi company page, with the Launch YC postRead in full on 27 October 2026, including the page data that dates the launch post to 10 November 2025. Source for the batch, the founders, the team size and the launch text.
- C Karumi home page, read 27 October 2026Read in full. Source for the product, the demo counter, the AI disclosure, the testimonial and the rating claims.
- C Karumi about pageRead in full. Source for the company's account of how it started, the open roles and the competitor comparison pages.
- C Karumi pricing pageRead in full, including the FAQ. Source for the plans, the absence of published prices and the CRM integrations.
- C Y Combinator job posting: Product Engineer (Full-Stack) at KarumiRead in full. Source for the web stack and the hosting on AWS.
- C Y Combinator job posting: AI/ML Harness Engineer at KarumiRead in full. Source for the voice, browser-agent and LLM parts of the product.
- C Karumi customer story: SalesforgeRead in full. Written by Karumi. Source for one customer's demo volume and trial conversion and for how the agent is trained.
- C Karumi Terms of Service, last updated 28 December 2025Read the opening sections. Source for the description of the service.
- D StartupHub.ai: Karumi profile (markdown version)Read in full on 27 October 2026. A startup directory; its funding and employee numbers are estimates. Cited only for the US$2 million funding total and the 10 employees.
- C Loot Drop database export, 1,749 failed startupsLocal export. AI-written summaries, not independently verified. Source for the failed predecessors.