Dmitry Pyanov

I'm building AI-native one-to-one learning experiences at LearnVector, a new education startup founded by Andrew Ng and backed by $100M from Coursera. I've spent the past three years in Andrew's ecosystem, at AI Fund and DeepLearning.AI, building AI Andrew Ng, a conversational replica of him for intellectual companionship. You can talk to it at avatar.andrewng.org.

I've been working on AI companionship since 2016, when I joined the Replika team, the year after the company came out of Y Combinator as Luka. Replika began as a memorial bot the team built for my friend Roman Mazurenko, who died in 2015. Strangers wrote in asking for bots of their own, and we pivoted Luka into Replika, the first AI friend. Roman was also the person who first told me about Andrew Ng, in 2012. Coursera had just launched, and Andrew had just published the cat paper, which showed that a neural network could learn to recognize cats from unlabeled YouTube videos. Andrew was the reason I got into AI, so working with him means a lot to me.

Open source, September 2026

Frog

The frog on this page is a desktop pet for the Mac, and an open-source alternative to Wispr Flow and Granola that runs entirely on device. Hold the fn key and talk, and it types what you say into whatever app you're in. Before a call, it starts taking notes: it records both sides, labels who said what, and writes a summary. Hold the right Option key and you can talk to it. It answers out loud, remembers what you tell it, and sets reminders. The models are local, so nothing you say leaves your Mac. The code is on GitHub under the MIT license. The one wandering around here is the same frog. Click it and it talks. You can pick it up, too.The one at the top of this page is the same frog. Tap it and it talks.

•••

Building now, since 2026

LearnVector

LearnVector is a new education company started by Andrew Ng to turn learning from one-to-many into one-to-one. Fifteen years after Coursera changed who could take a course, the course itself still teaches everyone the same way. LearnVector is building a personal learning guide: it starts by understanding you and your goal, lays out the way there, adapts the pace and the examples to you, and stays with you until you have actually mastered the skill, not just finished the course. It starts with an investment from Coursera and plans to work closely with Coursera and Udemy. The first product is coming in early 2027.

•••

Building now, since 2024

AI Andrew Ng

AI Andrew Ng is a conversational replica of Andrew, made for intellectual companionship. You can talk to it by voice or text. The first version launched at RealAvatar, where I led product. I have kept building it since at DeepLearning.AI, and a new version is live now at avatar.andrewng.org.

The work is to capture how Andrew comes across in a real conversation, which is harder than it sounds. A real person's voice has a lot baked into it — memory of who you are, empathy for what's important to you, knowing when to push back instead of just agreeing — and that's the work of orchestration: making the whole thing hold together as a conversation Andrew would say is his.

Andrew wrote about this in The Batch in May, calling AI Andrew "an AI companion shaped by my personality." On the build itself, he wrote: "I'd never before tried to codify this in an agentic workflow. This turned out to be hard and is still a work in progress."

AI Andrew Ng chat interface
•••

Earlier, from 2016

Replika

Replika became the first AI companion app to reach millions of users. This was 2017, before LLMs could hold a conversation on their own. The product had to feel like it remembered you and had a personality. The language models we used couldn't do either of those on their own — we built that infrastructure around them, by hand.

Roman Mazurenko was my roommate. He died in a car accident in Moscow in 2015. He had founded Stampsy, a digital storytelling startup, and was at the center of the Moscow creative scene that overlapped with Luka, the chatbot company I was at in San Francisco. Luka had come out of Y Combinator with what we used to call a swiss army knife AI platform — Silicon Valley HBO show characters, restaurant recommendations.

Roman Mazurenko in New York

Eugenia Kuyda, Luka's co-founder and one of Roman's closest friends, collected about eight thousand of his old text messages from his family and friends. She trained a neural net on them and put the bot online as a memorial. Strangers started writing in, asking if we could train one on someone they'd lost, or on themselves. So we pivoted. Eighteen months later we launched Replika to a 1.5-million-person waitlist.

Our approach was inspired by the Tamagotchi metaphor — a small AI companion you nurtured, that nurtured you back.

Some of what we shipped on a small team:

  • Long-term memory across sessions. Replika remembered facts about you and stayed in character.
  • "Train your Replika." Users shaped their bot's personality. Many spent hours making fictional characters, uploading hand-drawn avatars, building text games inside the chat.
  • Voice calls with five-to-ten second latency. We shipped it just before Alexa launched.
  • "Cake mode." Sending a 🍰 emoji switched the bot to a generative model fine-tuned for casual texting. Every reply got a thumbs-up or thumbs-down that we fed back into training. One of the earliest RLHF loops in a consumer product.
  • Replika Universe, where users could publish their custom Replikas for others to chat with.
  • Safety. We worked with mental health professionals to study how people were using the app and built classifiers for self-harm content that routed people to crisis hotlines.

Selected press:

•••
•••

Press