Correcto thu về 7 triệu USD để phát triển “Grammarly cho tiếng Tây Ban Nha”

The generative AI boom has put a spring in the step of Correcto, a Madrid-based language writing tool startup focused on Spanish speakers that’s today announcing $7 million in seed funding. The round is led by London-based Octopus Ventures, with Carya Venture Partners and River Park Ventures also contributing.

The founding team began work on their idea to build a Grammarly-style auto-editing tool for correcting written Spanish at the back end of 2021 — before generative AI tools like ChatGPT had blasted onto the scene and grabbed global attention.

Studying and working abroad led co-founders Abraham López Lee (CEO) and Ignacio Prieto Mayorga (COO) to use and appreciate tech tools like Grammarly which, they recount, helped them improve the quality of their written English. And they recount being surprised they couldn’t find comparable tools for correcting Spanish grammar and syntax, even as years of living outside Spain put a bit of a dent in their own confidence at writing professionally in their mother tongue, meaning they were all the more keen for a good tool to exist. So, along with a third (technical) co-founder, CTO Antonio Triguero Noriega, they set to work on an MVP.

The early version of Correcto used rules-based natural language processing plus a proprietary data-base of Spanish phrases to power Grammarly-style auto-editing features, correcting for grammar and style in the target language. The initial launch was as a Chrome extension to test demand. They’ve since also launched a freemium web app — and say they’ve seen 120,000 downloads (reporting some 70k active users) to date.

Over the period they’ve been building their product the competitive landscape for AI writing tools has of course shifted massively — thanks to the rise of widely available generative AI tools. Large language models (LLMs) like OpenAI’s ChatGPT and Anthropic’s Claude are now freely available (at least for casual) and able to generate all sorts of written texts on demand, including writing in Spanish. So LLMs could be seen as a threat to AI auto-editing tools. If you can just get a machine to produce the text for you why do you even need an auto-editor? Don’t those intermediary steps just get wiped away?

Correcto’s founders reckon not, of course. Firstly they argue the quality of Spanish texts produced by LLMs mostly trained on English language inputs is just not that great. Sure they can output Spanish. But López Lee suggests that, to a native speaker, the results sound a bit like how a child would write — which is suboptimal for the professional business users they’re most focused on. (The same conviction is driving another Spanish startup we covered recently, called Clibrain, which is working on building its own foundation model fine-tuned for Spanish language inputs.)

Secondly, the act of supporting people to improve the quality of their writing through a real-time auto-editing interface, which is what Correcto offers, sets up conditions for individual learning too via an interface that provides on-going feedback. Whereas, the LLM scenario — where a user just asks an AI to do the work for them and gets a new (or completely corrected) text automatically — doesn’t build in the same granular opportunities for human learning. This is why Correcto talks in terms of building a tool that can provide “augmented intelligence”, rather than artificial intelligence, as they say they want to keep the (human) user very actively in the loop.

“We do not necessarily focus on doing people’s jobs but really complementing and empowering people to write better and the core of our identity is augmented intelligence, not artificial intelligence. We want to help (our target users) across the world to get closer to their goals, feel more productive, feel more useful but at the same time not feel like they’re being replaced — or at least not pulling into the hole of, like, ‘AI will do everything for me’,” says López Lee, fleshing out their vision for a writing tool that, yes, boosts productivity but also preserves “the touch of personalization” — the “voice” and “cariño” — that can come through via someone’s writing style.

With the prospect of unending volumes of machine-generated text flying at human eyeballs, it doesn’t seem a huge stretch to imagine that comms which manage to maintain the human touch, and don’t edit out individual warmth and personality, could command an attention-grabbing premium vs the impersonal, spammy feeling you get from receiving something that sounds purely artificial. Including (and even especially) in a professional setting. And it’s selling a SaaS to enterprises that Correcto’s team is working towards. (Though they say they are also committed to keep offering a freemium version of the tool for individual users, which provides them with valuable feedback on the product.)

“It’s something that — we might be wrong — but we don’t see disappearing. At least in the next five to 10 years,” predicts López Lee of the personal touch conveyed through individual writing style. “And Correcto is here to stay and support cariño.”

This thesis goes some way to explaining why the team is bullish about the hyper competitive “AI boom” going on around them. Indeed, they suggest the viral success of ChatGPT et al has accelerated the market opportunity for their own product.

They have also directly tapped into OpenAI’s API to add a quasi-generative-writing feature called “Write for Me”, where the user selects the context for a text project they want help with, specifies the kind of audience and preferred tone of voice etc in order to get more than grammar and syntax corrections to assist them with producing an appropriate piece of writing in Spanish. So, basically, they’re leveraging mainstream generative AI developments to offer a richer form of auto-editing for Spanish — powered by their own Spanish language nuance “fine-tuning” of other LLMs.

“Right now we are in the process of fine tuning Claude from Anthropic and Davinci from OpenAI — because there’s an incredible gap when it comes to Large Language Models in English versus other languages,” argues López Lee.

“ChatGPT has opened us so many doors,” he goes on. “Because when we started Correcto — before ChatGPT existed — a writing tool wasn’t that much of an urgent necessity for corporations. Now it is. So now they’re interested. Now they’re hearing what you have to offer and how you can differentiate yourself with other offerings. And we’re in a good position to scale. So even though there’s competition, this competition is not — let’s say — 360. It doesn’t cover as many requirements of writing as we do. And the market is big enough to have a multiple unicorns — we strongly and deeply believe. But our goal is to be the first unicorn that comes from writing particularly focused in Spanish.”

The startup points to the huge size of the addressable market, saying there more native Spanish speakers (500M) than native English speakers (373M). Although that doesn’t factor in demand for English writing support from non-native English speakers — of which there are over a billion people with English as a second language (and there are far fewer second language Spanish speakers)… But playing second fiddle to English still sums to a lot of potential users. Albeit, if the tool’s output and utility vs LLMs is strongest for specific dialects of Spanish that implies a substantially smaller addressable pool than total native speakers.

Correcto’s co-founders see particular opportunity for their product in Latin America where native Spanish speakers may not have had access to a level of education that enables them to be confident about writing professionally in their mother tongue.

Per Prieto Mayorga, the top five countries for Correcto use currently are Colombia, Mexico, Argentina, Spain and the US. While initial target customers for paid versions of the product are professionals working in editorial, communications and marketing — aka “everyone who makes a living through writing”.

“At the end of the day, AI has exploded as a boon and so many companies are unsure on what’s the next step in order to adopt AI,” López Lee suggests, sketching out that their hope, as a plucky startup, is to ride the hype wave and be ready to offer the expected flood of Spanish businesses suddenly looking for AI writing support with a “one stop shop” solution.

Longer term, he suggests the AI boom could also generate (ha!) the possibility of a strategic exit for Correcto to one of the AI giants — assuming the results of its fine-tuning of their models wins it enough Spanish speaking professionals to scale usage to a critical mass that lands it on the radar of the likes of OpenAI.

But couldn’t such well resourced AI giants just do this kind of fine-tuning work themselves to up their LLM game in non-English languages like Spanish?

Correcto’s team doesn’t think so. “We are really focusing on new instances of the Spanish language — like dialects — that are currently untouched and data gathering for these particular dialects are incredibly complex and inaccessible,” argues López Lee. “I’d say this puts us in an interesting exit strategy. When we do this in-house — if we build incredible technology, incredible databases, a really strong fine tune of these models that are particularly crafted towards Spanish and its dialects — we don’t see why this wouldn’t be an incredible exit opportunity for us in the future from one of these large players.”

That possibility is also why the team is focusing its efforts on building out a compelling product.

“(AI giants) are far ahead in terms of resources and human capital resources and the quality of that human capital is incredibly demanding and hard to find — and we wouldn’t be able to compete from a Large Language Model perspective with these companies. Where (we want to) compete and add value to societies (with) our products (is by) growing and improving towards what our users want. So… we’re more focused on giving people a solution versus competing with these Large Language Models, because we don’t see any point on that as they’re incredibly strong.”

Elsewhere, as regards competition, there is the aforementioned Clibrain — another Spain-based startup that’s fully focused on tuning AI to serve Spanish speakers. López Lee argues their respective use-cases are distinct (at least for now), with Clibrain gunning to build its own Spanish-language tuned LLM to sell API access to others, while Correcto is “much more focused on productizing our Large Language Model fine tunes and selling that to companies in the form of a product, not in an API”. He even floats the possibility that the pair could end up working together in the future — if/when Clibrain delivers a more powerful LLM (with 40BN parameters, rather than the 7BN of its debut model).

On the auto-editing side, he also name-checks rival writing assistant business LanguageTool. Although it’s not solely focused on Spanish — offering grammar help across multiple languages — which is one obvious difference.

For language generation, he admits there’s a huge amount of competition out there — not least from mainstream LLMs like ChatGPT (which can certainly generate Spanish language copy that’s good enough for many people), to smaller tools targeting help at Spanish speakers with a specific content focus, such as writing emails or SEO copy. Correcto’s strategy is to go deep on language nuance and broad in the sense of offering a tool for Spanish speakers that offers multifaceted utility.

Prior to today’s $7M seed the startup had taken in a pre-seed raise of $1M. At a crucial earlier moment, when the team was kicking around the idea of building an MVP, they also recount being buoyed by winning an entrepreneurship prize awarded by King’s College Cambridge — which gave them £20k to start building. “That was important for us back then,” recalls López Lee. “We weren’t students but we didn’t have any financial means. It was impossible for us to have £20,000 or even £3k and that was a huge (moment).”

The seed funding they’ve announced now is of course rather more sizeable — indeed, their PR bills it as one of the highest rounds to date awarded to a Spanish startup. This new cash injection is being ploughed into AI and product development as Correcto works on sharpening a differentiating edge, including by improving its AI tech’s ability to set style and tone and style in Mexican Spanish, Colombian Spanish and Argentinian Spanish.

Commenting on the seed in a statement, Andrés Pérez Soderi, general partner at Carya Venture Partners, also highlighted the addressable opportunity in LatAm: “Correcto signifies more than just grammar correction; it brings generative AI, cultural connection and opportunity to millions of users in LatAm. We could not be more excited to back this mission and its driven team.”