Building AnyLinga: 3 prototypes in 60 days from Argentina

It was almost 2 a.m. in Buenos Aires when I realised I was the bottleneck in my own sales pipeline.

I’d moved to Argentina at the start of 2026 to get closer to the Latin American market. I’d spent the previous several years building integrations and automations on top of amoCRM and Kommo, mostly for Russian-speaking clients. The plan was simple: open a fresh Kommo pipeline, run a couple of outbound campaigns to LatAm leads, learn the market from the ground.

One small problem. I didn’t speak Spanish.

This is the 60-day log of how AnyLinga went from that 2 a.m. realisation to a Kommo Marketplace widget translating ~300 conversations a day. It’s also a defence of building a B2B SaaS as a solo founder in 2026, AI assistance and all.

---

Why did I build this? The 2 a.m. Buenos Aires problem

Three out of four consumers won’t buy from a business that can’t communicate in their native language — 76% of consumers prefer to buy from companies that speak their language, with 40% saying they simply won’t purchase otherwise (CSA Research, 2020 B2C study, 8,709 consumers across 29 countries). I knew that statistic before I moved. I underestimated what it actually felt like to be on the other side of it.

My first week in Argentina, my workflow looked like this: open the Kommo conversation, copy the customer’s Spanish message, paste it into Google Translate in another tab, read the English version, type my reply in English, paste that into DeepL or Google Translate one more time, copy the Spanish translation, paste it back into Kommo, send. Roughly 30 to 120 seconds per message, depending on length.

By the third or fourth conversation it was obvious this wouldn’t scale even to a single person. The reply time was minutes, not seconds. I made copy-paste mistakes — twice I almost pasted a client’s name into the translate window. And there was no way to hand a conversation to anyone else later, because the translated text never lived inside the CRM.

I checked the market. Every “WhatsApp translator” I could find was a browser extension that worked on WhatsApp Web, outside of any CRM. Nothing connected to Kommo. So in true founder fashion, I told myself I’d build a quick weekend prototype just to get me through Monday.

---

Prototype 1 — a Telegram bot in a weekend

The first version was the smallest thing I could ship that would unblock myself: a Telegram bot. I’d forward a customer message to the bot, the bot would translate it through DeepL, send the translation back to me as a reply, and I’d copy it into Kommo. Round-trip, maybe 8 seconds.

It took roughly 48 hours to ship, mostly because I wanted it to handle attachments and voice notes too. The transcription was rough but usable. By Monday morning, I had a working tool I could use during a sales call.

What it proved:

• DeepL plus OpenAI for fallback was good enough translation quality for sales conversations — even on slang and regional Argentine Spanish.
• Latency wasn’t the constraint; the manual steps were.
• Telegram’s API was a fine sandbox for the V1, but it was clearly the wrong end-state — I shouldn’t be the bridge between two apps.

What it didn’t fix:

• I was still copy-pasting between Kommo and Telegram. Four manual actions per inbound message.
• The translated text wasn’t logged in the lead card — only the original Spanish was. So if I scrolled back two weeks to remember what a customer had said, I was reading Spanish again.
• I couldn’t hand any of this to a sales assistant. The bot was bound to my own Telegram account.

The prototype proved the translation engine. It didn’t prove the workflow. So I started planning V2 the same week.

---

Prototype 2 — listening to Kommo directly

Version 2 was the obvious next step: cut myself out of the relay loop. I built a small service that subscribed to Kommo webhooks, translated incoming WhatsApp messages on arrival, and pushed the translation back into the same conversation as a CRM note. The agent — me — could read the English version directly inside the Kommo card.

The latency dropped from 8 seconds (with my Telegram round-trip) to under 2 seconds end-to-end. Translation appeared in the conversation panel a moment after the original Spanish message arrived. It felt like a different product.

But V2 was a half-solved problem. Inbound worked. Outbound didn’t. I still wrote replies in English, and the customer received them in English. Most of my Argentine leads stopped responding after my second message.

The lesson from V2 was uncomfortable but useful. Half-automating a two-way conversation is a different kind of broken than not automating it at all. With the manual copy-paste workflow, at least my replies were slow but in the customer’s language. With V2, my replies were fast but in the wrong language. The customer experience got worse before it got better.

So V2 ran for maybe two weeks. Then I shut it down, threw away the partial outbound code, and started V3 properly.

---

Prototype 3 — the full bidirectional widget (this is what’s live today)

V3 is what you can install today as AnyLinga from the Kommo Marketplace. It translates in both directions, surfaces inside Kommo’s own conversation panel as a native widget, and logs both the original and translated text into the lead record. The agent writes in their language, the customer reads in theirs, and the history stays clean for whoever opens that lead three months later.

The build from prototype 2 to a public widget took about 40 of the 60 days. The Kommo Marketplace OAuth flow alone ate a week, mostly because Kommo’s authorisation surface for public integrations has some legacy quirks that aren’t well documented. The Meta side — getting AnyLinga registered as a WhatsApp Business API integration partner — took longer than the entire V1 prototype.

Beyond Kommo and WhatsApp, V3 picked up Telegram, Instagram Direct, Facebook Messenger, and TikTok DMs along the way. Each channel has a slightly different API shape; the abstraction layer to handle them all uniformly is honestly the part of the codebase I’m proudest of.

The reason those four pairs lead is simple: they’re the corridor I personally live in. English ↔ Spanish is the LatAm-to-US bridge. Spanish ↔ Russian is the legacy audience inherited from years of Kommo work in CIS markets. Spanish ↔ Portuguese is the Argentina-to-Brazil crossover. Portuguese ↔ English is the Brazil-to-US bridge that’s growing fastest right now.

One thing worth saying clearly: I don’t have many paying customers yet. The product was on Kommo Marketplace and accepting trials starting late April 2026. The traffic you see in those numbers is mostly trial accounts plus my own selling motion. That’s honest. The traction is real but small, and the goal of the next six months is to make those numbers ten times bigger.

---

What did I learn vibe-coding a B2B SaaS in 60 days?

The unfashionable confession: significant portions of AnyLinga are AI-coded. Not in the “ChatGPT wrote my product” sense. In the “Claude pair-programmed every prototype, generated the Kommo widget skeleton, and shipped the locale boilerplate” sense. I won’t pretend a non-existent CTO did it.

What AI actually accelerated:

• Routine code — CRUD endpoints, schema migrations, locale string interpolation, the Kommo widget’s JS skeleton.
• Translation between conventions — converting an API spec into a typed client, or a webhook payload into a normalised internal event shape.
• Documentation drafts — the first pass of the knowledge base and this blog were both AI-drafted, then heavily edited.

What AI did not help with:

• Product decisions. Should AnyLinga support TikTok DMs in V3 or wait? Should Spanish regional variants (LatAm vs Castilian) be auto-detected per contact or set per pipeline? Those are taste calls, made by the operator-founder — me.
• Legacy quirks of platform APIs. Kommo’s OAuth flow, Meta’s business verification, amojo channel registration — AI hallucinated plausible-looking solutions that didn’t match reality. I spent the most painful hours reading docs and trial-and-erroring those by hand.
• Customer empathy. The decision to throw away half-built V2 outbound code came from listening to actual customer reactions, not from any IDE.

The pattern that worked: human as operator, customer, and designer; AI as the typing. If I’d hired a junior developer in week one, I’d still be onboarding them now. The fact that a single operator can ship a multi-channel B2B SaaS in 60 days isn’t a flex. It’s just what the 2026 stack looks like, if you’re willing to own that out loud.

---

What’s next?

Three things on the immediate roadmap:

Locale parity. AnyLinga’s landing pages and knowledge base are already in six languages. The blog catches up over the next quarter — including this post, which will be hand-translated to Russian and Spanish before the auto-translate pipeline runs the rest.

Verticals. WhatsApp CRM in Latin America is dominated by real estate, travel, and B2B SaaS sales from the US into LatAm. Three vertical guides are scheduled for the next six weeks, with workflow specifics for each.

WhatsApp’s native translation, again. Meta launched in-app message translation in September 2025. It runs on the mobile WhatsApp app only and doesn’t integrate with any CRM. That gap is the entire reason AnyLinga has a market. I’ll keep tracking it.

If you’ve been hitting the same wall I did — selling into a language you don’t speak fluently, copy-pasting your way through every WhatsApp thread — I’d genuinely like to hear from you. The DM is open. So is the 7-day trial.

---

FAQ

Who built AnyLinga?

AnyLinga was built by Max Shishkin in early 2026. Solo founder, AI-assisted. The company is a product of VentaBoost LLC, registered in Delaware, United States. Max moved to Argentina in January 2026 to sell into Latin America and built the product because he needed it himself.

Why did AnyLinga start with WhatsApp and Kommo specifically?

Because the founder lived inside that exact workflow. Max had spent years building integrations on top of amoCRM and Kommo for Russian-speaking markets. When he moved to Argentina, his daily reality was answering WhatsApp leads through Kommo in a language he didn’t speak. The product had a first user — him — from day one.

What does “vibe-coding” mean here?

AI-assisted development where most boilerplate and routine code is generated and edited collaboratively with an AI pair-programmer, while product decisions stay human. In 2026 this is how many small operators ship at speed against bigger teams. AnyLinga’s backend, Kommo widget, admin UI, and integrations were all built with significant AI assistance over a 60-day window.

How many conversations does AnyLinga translate per day right now?

Live infrastructure currently handles roughly 300 conversations and 1,000 messages a day across four primary language pairs: English ↔ Spanish, Spanish ↔ Russian, Spanish ↔ Portuguese, and Portuguese ↔ English. The product translates 50+ languages, but those four are where production traffic concentrates.

---

Conclusion

Three prototypes, sixty days, one operator-founder-customer in a Buenos Aires apartment who didn’t speak Spanish on the day he arrived. The interesting part isn’t that AnyLinga exists. The interesting part is the build pattern: ship a hacky V1 to one user (yourself), let that user’s friction tell you exactly what V2 should be, repeat.

If you want the systematic, four-method comparison of how to translate WhatsApp inside any CRM, that’s the next read: Translate WhatsApp in CRM: 4 Methods Compared. If you’re thinking about how this scales beyond a solo founder — what international sales teams actually do with this kind of layer — that’s WhatsApp CRM for International Teams.

Or, if you’d rather just see the thing translating live: install it from the 7-day trial. No credit card. I’ll be on the other side of the DM.