Case studies in numbers, the way a report does it

I decide what fits me.

A pilot concept for an existing denim brand. Product fit becomes evidence for a broader cultural idea: the woman chooses clothes for her own body and life. The brand offers a reusable language of personal choice.

Client brief: Deepen the meaning of SLAD and connect clothing, freedom of choice and customer participation in one brand system.

Pilot concept; no launch or performance results are claimed.

Skin changes. Care can change too.

A skincare brand concept for ages 15–18. The brand focuses on how skin feels today and provides a simple base routine and an open language. The colon lets the audience complete the name with their own status.

Client brief: Create a youth skincare brand concept from scratch, from positioning and naming to product logic, visual language and participation mechanics.

Pilot concept; no launch or performance results are claimed.

Advertising holding a 10.2% cost of sale against a 15% target

An online store selling professional cosmetics. Campaign management and audit: clearing out dead keywords, switching off junk display placements and moving budget into the strong segments — abandoned carts and past buyers.

Best segments: carts at 4.3% cost of sale, past buyers at 3.5%. The audit found a further ≈₽43,000 a month in losses.

A site search engines couldn’t see, and display traffic burning the budget

Professional hair care. The SEO audit showed that 65% of the links in the sitemap returned 404 — for months, crawlers had been walking pages that no longer existed. A parallel review of the ad account found budget draining into junk display inventory. Both reports came with a 30-60-90 plan.

The full loop: from front-end markup to offline conversions

A fitness studio in Saint Petersburg. We finished the website (speed, contrast, calls to action, responsive images), launched Yandex Direct from scratch, connected online payments with SMS notifications, and set up offline conversion upload into analytics — so automated bidding learns from real sales rather than clicks.

The proving ground where we test everything before clients see it

Our own school teaching AI tools runs on the same stack we sell: webinar funnels in the LMS, traffic from paid search, follow-up sequences and a dashboard where revenue is reconciled against the bank and net profit is calculated by formula rather than by feel.

net = (revenue − payment fees) × 0.88 — the formula from our own dashboard.

AI agents already working with live customers

Our content generation service is run by agents: the support agent answers customers with drafts approved over Telegram, an SEO agent built dozens of landing pages and a 50-lesson course, and a marketing agent optimises the ad account every day. All of it is process we then transfer to clients.

The site took 46 seconds to load — and one line kept it out of search

A full-day private school in St Petersburg with three campuses. The audit scored it 3.3 out of 10: robots.txt blocked almost the entire site, the sitemap listed two URLs instead of the real structure, the personal-data consent link under every form returned a 404, and the mobile homepage carried an uncompressed 8.3 MB image. In the same run we did not just document the problems — we rebuilt the site against that list: 29 static pages instead of WordPress on PHP 7.1, the mandatory «About the educational organisation» section with 147 documents, and a low-vision mode.

LCP measured with Lighthouse on a mobile profile, slow 4G and 4× CPU throttling. Demo version: bwschool.demo.socode.ru

Beauty brand visuals: 5 WB/Ozon listing images from one photo

From a single photo of the bottle and a short product description, our API assembles a set of listing images in one brand style: a cover, a benefits infographic, specifications, an in-use lifestyle shot and a close-up. The bottle, label and logo look exactly as in the source photo on every frame — the model does not redraw the product. The Russian copy on the images is taken word for word from the product data: no “clinically proven”, no prices, no promises that aren’t in the description.

The client writes no prompts — only the name, benefits and specifications. A cream gets its own specs, a shampoo its own: the template inserts exactly what was sent, and a frame with no data is skipped rather than filled with invented text. That is why the same set scales to a whole catalogue — through the Genosai API or an MCP agent in Claude or Cursor that works through the products on its own.

The LUMA brand and product are for demonstration; the bottle photo is generated too, to show the pipeline at work. In a client project the input is the client’s own photo and product listing.

From an expert's notes to a finished course with cartoons in 5 days

An expert in children's financial literacy came to us to launch online learning for children aged 3–17 and their parents. A team of seven AI agents — instructional designer, speechwriter, presentation agent, designer, video director, developer and reviewer — built the programmes with lesson-by-lesson scripts, a brand identity with a beaver-cub mascot, five educational cartoons, cards with QR audio and the course website. Every stage went ahead only after the expert approved it.

Full case study →

Its own CRM instead of a boxed service — and an agent that improves it on the team's request

A follow-up to CASE 03. The studio was leaving a cloud booking service: it needed its own payroll schemes, room limits, a multi-service visit on one receipt and client data in its own database. From the owner's 21-minute video the agent produced a 67-item spec, and we built a CRM the team found familiar. Then we added a Change Center to the admin panel: the owner and front desk write a task with a screenshot, the developer agent makes the change in minutes, and the team accepts it or sends it back.

Full case and task examples →

The task was dictated by voice — in one evening the agents delivered a new website, catalog photos, a lookbook and a selling carousel

A premium denim brand from St. Petersburg. The old website sladbrand.ru did not sell: the SEO audit scored it 3.3 out of 10. An entrepreneur with no AI experience dictated the task in three voice sentences. Gena, the orchestrator agent, translated it into prompts, estimated the cost and assembled a crew of eight agents on his own.

In one evening: a new selling website, 33 catalog photos in a single brand style, a 10-shot lookbook with no studio, model or photographer, a carousel for Instagram and VK, a 24-page SEO audit and a commercial proposal.

Open the new website → SEO audit of the old site — PDF, 24 pages, in Russian ↓

A Google AI Studio prototype became a product in one day: a website, login and 16 voice agents and coaches

The owner had built a voice assistant prototype in Google AI Studio and dictated the task: take the code, connect the new Gemini model, launch it on his own domain, create a bold landing page with a login and role settings — name, description, script, voice and knowledge base. The developer agent pulled the source straight from the browser, rewrote the server side, deployed the app on a server in the Netherlands (the Gemini API does not serve Russian IPs) and pointed the subdomain in the hosting panel itself.

In parallel, a research agent studied the market for voice agents and AI coaches and assembled 10 roles with instructions and knowledge bases, while a knowledge agent packed Anton's AI-agents course materials into a voice expert. Every role speaks first, answers from its own knowledge base and, after the call, saves the audio recording, transcript and a summary with action items. Coaches give a score and name areas for growth.

Open the Dobby website →