About this interview

The first episode of AI in Production, recorded at the AI Start Academy location in San Francisco. Lex Mustafin sits down with Pavel (Paul) Shaburov, founder and CEO of Glam.AI — a consumer, AI-content-first social platform.

Pavel walks through the full arc of building a consumer generative AI startup: five failed app attempts, finding product-market fit through unit economics rather than intuition, scaling to 16 million downloads and roughly $60M ARR in under two years, and building proprietary adapter models for identity preservation on top of foundational models. He closes with a concrete prediction about where AI-generated social media content goes next.

What you will take away

• Identify the two metrics that actually signal consumer product-market fit — revenue and retention — and why paying users are the only feedback that counts.
• Use paid performance marketing (Meta, TikTok) as a cheap validation instrument: scale ad spend and watch whether LTV:CAC holds.
• Explain what adapter models are, and why a fine-tuned layer on a foundational model beats the foundational model for a narrow niche.
• Understand identity preservation — and print/logotype preservation for merchants — as a real technical moat rather than a wrapper.
• Reason about the cost wall of cloud-based AI generation and why on-device (edge) compute is the likely answer.
• Apply iterative startup methodology: rapid pivoting, and hackathon-based recruiting of technical talent.

Topics covered

• Background and the path to Glam.AI — growing up in the US, computer science in St. Petersburg, companies acquired by Snapchat, then deciding to build his own consumer AI company.
• Five failures before product-market fit — iterating through five mobile app concepts, and what made the sixth different.
• Validating with paid performance marketing — on the failed attempts CAC always exceeded LTV; with Glam the ratio held as budget scaled, and then virality delivered users at zero cost.
• The product — a large filter/template gallery, no prompting required: upload a photo, get Instagram-worthy content in one to two minutes.
• Scale — 16M downloads, 2.5M monthly actives, ~7M pieces of content per month, ~500M views on content generated in a single month, 92% female user base.
• Team and hiring — 40 full-time, roughly half technical, office-based in Georgia (the country); co-founders and engineers recruited by sponsoring hackathons in Eastern Europe.
• Adapter models and identity preservation — why foundational models lose your hair colour, eyes and facial structure, and how a fine-tuned adapter layer keeps the output actually you.
• The prosumer and merchant use case — foundational models destroy printed logotypes and text; adapters preserve them, which turns product photography into a commodity.
• Where Gen AI is heading — Pavel’s argument that the act of creation disappears and social media becomes pre-generated, personalised consumption — and why that content stops reading as “AI slop.”

Key concepts

Product-market fit validation via paid performance marketing · LTV:CAC ratio · consumer generative AI applications · adapter models and fine-tuning on foundational models · identity preservation in AI image generation · on-device / edge compute · virality as zero-cost acquisition · iterative startup methodology · AI-generated social media content · prosumer and merchant use cases · ARR scaling in consumer AI · hackathon-based technical recruitment