Available for new projects · We build our own products

AI and automation for business
3–5× faster than hiring a team

A small studio that ships finished products in weeks, not months. AI agents, Telegram bots, websites, business automation — what large teams deliver in six months, we deliver in 2–4 weeks. Speed comes from a properly set up AI development process. Experience comes from the fact that every person on the team has shipped their own product.

5.7M+
On-chain transactions
(#1 on NEAR Protocol)
13+
Production AI & Web3
products shipped
10+
AI providers in production
OpenAI · Anthropic · Replicate · ElevenLabs · Kling · Luma · …
7
people in the team,
each shipped their own product
// Why teams choose us

Hiring takes months.
We ship in weeks.

No hiring cycle, no onboarding, no long-term commitment. You bring a scoped problem — we deliver a working product, documented for handover.

AI-accelerated delivery

Claude is the core of our workflow. One engineer with the right AI flow ships what used to take a team of 5.

🎯

We own what we build

Each of us builds and runs a product in our own ecosystem — so we treat client work with the same accountability.

🛠

Full-stack, not just vibe-coding

Speed comes from vibe-coding, but underneath sits classical full-stack engineering — we understand databases, load behavior, security, the code itself. Two-environment CI/CD, healthchecks, rollback playbooks, documented handover. Not prototypes — production-grade products that survive when the user count grows.

🧠

End-to-end ownership

System design, AI pipelines, infra, billing, mobile builds, SEO, content automation — we cover the whole stack.

// What we build

Six directions —
one full-cycle team

From AI agents and Telegram products to websites and business automation. We take a project end-to-end — from a 30-minute call to a deployed product — or step in on a specific scope to reinforce your team.

AI implementation

  • Consultant bot with access to your knowledge base — answers in the company's voice, handles typical customer support
  • AI agent with memory — remembers the conversation and your documents, doesn't mix up clients
  • Content factory — from one brief: script → voice → video → auto-publishing
  • Founder's voice in a bot — replies as the founder would, in their style
  • AI failover — one provider goes down, automatically switches to another
  • Cost control — hybrid of free and paid models tuned to the task

Web3 and blockchain

  • Custom NFT collection launches — from art to live mint, NEAR / TON
  • NFT marketplaces and on-chain auctions
  • Mass NFT distribution — 10,000+ wallets in a single campaign
  • Smart contracts on NEAR, TON, EVM and Solana
  • Token launches and tokenomics design
  • Crypto payment integrations and on-chain wallets

Telegram bots and Mini Apps

  • Sales funnel bot — qualifies leads, hands warm ones over to a manager
  • Mini Apps — a full web app inside Telegram with payments and personal accounts
  • CRM integration — bot writes to GetCourse, AmoCRM, Bitrix or a custom system
  • Live-stream automation — fair raffles, real-time entries, results in chat
  • Admin panel for non-technical staff — managers manage without touching code
  • Payments via YuKassa and crypto, referrals with anti-fraud

Websites and web apps

  • Landing pages in a day — fast hypothesis testing for marketing
  • Multi-page sites with SEO under specific search queries
  • Web applications on React / Next.js — SaaS, dashboards, admin panels
  • Programmatic SEO — auto-generate hundreds of pages from one template
  • Admin panels and back-office tools for daily team operations
  • Wallet auth flows and Telegram login for Web3 / TG projects

Automation and integrations

  • Connect services through APIs — CRM, payments, accounting, analytics in one circuit
  • Data parsing — collect prices, reviews, contacts from public sources legally
  • Database engineering on PostgreSQL / Supabase — clean structure under load
  • Telegram notifications and alerts — about a sale, an outage, a key event
  • CI/CD pipelines — auto-deploy on git push, no scrambling on release day
  • Scheduled tasks — reports, mailings, syncs run automatically

Mobile (Capacitor)

  • Native iOS and Android from one codebase — cheaper and faster than two separate apps
  • App Store and Play Store submission — we know the moderation traps
  • Push notifications and deep links — bringing users back to the app
  • Payments in regulated markets — Apple/Google Pay and local providers
  • Native bridges for hardware (camera, GPS, biometrics) — wherever the web isn't enough
  • Migrating an existing web app to mobile in 2–3 weeks instead of writing native
// Pricing

Popular services & starting prices

Our most-requested directions. Final price depends on scope.

AI implementation

AI agents, RAG, LLM chatbots, AI-driven process automation

from $1,500

Web3 & blockchain

Smart contracts, NFT infrastructure, multichain integrations, token launches

from $1,800

Web development

Landing pages, websites, web apps, SaaS dashboards, admin panels

from $1,000

Telegram bots & Mini Apps

Bots with logic, payments, CRM integration, Mini Apps

from $800

Automation & integrations

Connecting services via API, CRM, notifications, CI/CD, cron jobs

from $700

Mobile (Capacitor)

Native iOS & Android via Capacitor, single codebase, store submissions

from $900

Final price is set after a free discovery call — it depends on scope and complexity.

// Selected work

13 production products,
shipped end-to-end

A mix of confidential client work (under NDA) and products from our own Web3 ecosystem. Each case is real, in production, with measurable outcomes.

Web3 · Flagship

Yupland — Web3 Ecosystem

NEAR · TON · React · Node.js
Details →

Yupland — Web3 Ecosystem

Web3 · Flagship
Role
Full-cycle Web3
Timeline
1.5 years live
Chains
NEAR + TON
Funding
Self-funded
cases/yupland_title.jpeg

A self-built, multi-chain Web3 ecosystem on NEAR — 15+ products linked by one economy and 21 tokens, where real users have lived, traded and crafted for two years across two market cycles.

The challenge

Web3 games die with the market — in a bear, NFTs rot in wallets, communities scatter and metrics are bot-inflated. The goal: an ecosystem whose on-chain activity doesn't depend on price, driven by real wallets, not bots.

Screenshots: YupPortal — the ecosystem's hub & navigator web app (marketplace · DEX · incubator)
What we built

The Reputation Game — an antifragile loop of trade → craft → burn → win that gets stronger under pressure. Cross-chain crafting (Alchemy Lab), a burn wallet (Duplo), mass NFT distribution (Sender — 120 NFTs per tx) and AI products, all live on NEAR + TON on custom indexers and smart contracts.

Alchemy Lab

Cross-chain NFT crafting across NEAR + TON — 29K+ crafts, 283 recipes. Burning solves inflation; recipes drive retention.

NFT-Sender

Mass multi-chain distribution at 120 NFTs per transaction, past the NEAR throughput limit — no analogs on the market.

YupSoul — AI Music Oracle

The flagship AI product: personal songs from a natal chart across web, Telegram, VK, iOS & Android in 5 languages.

5.7M+ on-chain transactions, 123K+ holders, 21 tokens — all self-built, on a chain that ignored us.
Under the hood
  • Custom NEAR indexers tracking every on-chain event in real time
  • Sender pushing 120 NFTs per transaction past the NEAR throughput limit
  • Cross-chain bridge logic uniting NEAR + TON liquidity in one loop
  • Self-sustaining for 1.5 years — zero grants, zero VC
Results
5.7M+ on-chain tx #1 by activity on NEAR 123K+ holders 190K+ NFTs 15+ products · 21 tokens $3K BitGet win
Stack: NEAR/TON smart contracts · React · Next.js · Node.js · TypeScript · Python · Telegram Bot API · custom indexers
Web3 · Portal

YupLand Portal — Ecosystem Hub

React · TS · Vite · NEAR + TON
Details →

YupLand Portal — Ecosystem Hub

Web3 · Portal
Role
Full-stack
Platform
Web + Telegram Mini App
Built by
1 engineer
Live since
Oct 2025
cases/portal_title.jpeg

YupLand Portal is the navigation layer of the Yupland ecosystem — one app, web + Telegram Mini App, that bundles a custom CMS, DEX, incubator, wiki, gamification layer (levels / achievements / quests / loot boxes / NFT staking), YupPay, prediction market (Polymarket-style), YupDrop, subscriptions, referrals, social graph and a Telegram notification engine. Built solo by one engineer.

The challenge

The Yupland ecosystem has 15+ products and 21 tokens spread across NEAR + TON. Users needed one place to navigate the whole world — wallet, balances, staking, games, marketplace, governance — without bouncing between apps. The portal also had to host games and services built by the community and incubator projects, not just by the studio.

What we built

A single full-stack app — web + Telegram Mini App on one codebase — wrapping the entire YupLand world. Custom CMS, DEX trading every ecosystem token pair (13 NEAR FTs + 8 TON jettons), incubator with no-strings community funding, gamification layer (levels, achievements, quests, daily loot boxes, NFT staking), YupPay, YupDrop, Polymarket-style prediction market, subscriptions in Darai, in-portal transactions, friends graph, events calendar, promo codes, reward pool and a Telegram notification engine for users and admins. Vibe-coded end-to-end.

Wallet & DEX

Connect a wallet, see your Yupland tokens live on-chain, watch the Darai chart. Built-in DEX trades every ecosystem pair across 13 NEAR FTs and 8 TON jettons — from MED to Darai to bees.

Gamification + Energy economy

Levels, achievements, quests and the daily Duplo loot loop — ~2M boxes opened per month (rewards batched into single transactions to save NEAR gas). NFT staking earns Energy, spent on the same loot loop.

Games marketplace (hybrid build)

10+ mini-games in one portal — Flappy, 2048+, Reputation Calculator, YupLottery, YupLink, feeding checkers, plus deep links into YupSoul, YupDar, Sender and KeksVPN. Built by the studio, by the community and via the incubator.

Incubator (no-strings crowdfunding)

Project applies → admin approves → community funds it freely in NEAR / USDT / Energy / Darai. 9 projects live. Some become standalone ecosystem products.

~2M loot boxes opened in 30 days, a DEX trading every token in the ecosystem, 9 funded projects in the incubator — all from a Web + TMA app built solo by one engineer.
Under the hood
  • Web and Telegram Mini App share a single codebase
  • Custom indexer mirrors YupAI NFTs from the wallet without moving them
  • Loot-reward batching — multiple rewards stacked into one transaction to save NEAR gas
  • Throughout: workarounds for NEAR API rate limits and throughput constraints
  • Telegram notification engine — rewards for users, low-pool alerts for admins
Results (last 30 days)
~2M loot boxes / 30d MAU ~200 ~700 wallets connected 9 projects in incubator Built solo · vibe-coded
Stack: React · TypeScript · Vite · Node.js · PostgreSQL · FastAPI · Realtime · Edge Functions · Tailwind + shadcn/ui · NEAR + TON SDKs · Cursor + Claude Code (vibe-coding)
AI · Mobile

AI Music Generation Platform

Telegram · Capacitor 8 · Node.js · LLM
Details →

AI Music Generation Platform

AI · Mobile
Role
AI engineering
Platforms
TG Mini App + iOS & Android
Type
Client product
Backend
1 unified
cases/ai-music_title.jpeg

An end-to-end AI music platform delivered across three surfaces — a Telegram Mini App and native iOS & Android apps — all on one unified backend.

The challenge

Generate music reliably at scale and accept payments in a regulated market — all from a single codebase that ships to web and both app stores.

What we built

A multi-stage AI pipeline with retry/failover, one backend powering all three clients, three payment integrations including card acquiring in a regulated market, and two-environment CI/CD with Sentry monitoring.

Results
3 platforms 1 backend 3 payment integrations Two-env CI/CD
Stack: Node.js · Telegram Bot SDK · Express · Supabase · Capacitor 8 · LLM provider · Music gen API · Sentry
AI · RAG

AI Sales Funnel + RAG Agent

grammY · pgvector · DeepSeek · React
Details →

AI Sales Funnel + RAG Agent

AI · RAG
Role
AI + full-stack
Platform
Web + Telegram
Type
Client (online school)
Timeline
6 weeks

A complete AI-powered sales funnel for an online education business — from landing pages to a RAG agent that answers in the founder's own voice.

The challenge

Turn a founder's huge content library into an always-on sales agent that genuinely sounds like them — wired into the existing CRM and protected from referral fraud.

What we built

Two landing pages, a multi-step Telegram onboarding bot, a RAG agent trained on the full content library, GetCourse CRM integration, an anti-fraud referral system and an admin dashboard — delivered in 6 weeks.

Results
6 weeks delivery Founder's voice RAG GetCourse CRM Anti-fraud referrals
Stack: Node.js · grammY · PostgreSQL/pgvector · Voyage AI · DeepSeek · Express · React · GetCourse API
Web3 · Infra

NFT-Sender — Multichain NFT Infrastructure

NEAR · TON · EVM · Solana · Rust
Details →

NFT-Sender — Multichain NFT Infrastructure

Web3 · Infra
Role
Web3 infrastructure
Throughput
Up to 120 NFTs / tx
Chains
NEAR · TON · EVM · Solana
Status
#1 on NEAR by tx
cases/nft-sender_title.jpeg

NFT-Sender is the multichain NFT infrastructure powering bulk distribution, marketplace, on-chain auctions, an explorer and a reputation game across NEAR, TON, EVM and Solana. Currently #1 on NEAR by transaction volume — 1M+ tx on a single collection, 40 collections live, 10K+ wallets in the reputation system. Built from scratch — no analogs on the market.

The challenge

Mass NFT distribution on NEAR runs straight into the chain's throughput limit — a few NFTs per transaction means days to airdrop a 10K-wallet campaign. Beyond raw distribution, an NFT ecosystem also needs the layers around it: a marketplace, auctions, a wallet-level explorer, and a reason for holders to actually engage with what they own. Ideally it should reach across chains, not just one.

What we built

A custom infrastructure that pushes up to 120 NFTs per transaction on NEAR — a workaround for the chain's throughput limit, with no analogs on the market. Around the distribution sit a built-in marketplace and on-chain auctions, an explorer that parses 10K+ wallet addresses with full on-chain data extraction, and a reputation game where each NFT carries a positive or negative reputation coefficient and wallets compete for TOP+ and TOP- rankings. Alchemy Lab — the cross-chain crafting mechanic — plugs in here too. Cross-chain by design: NEAR, TON, EVM and Solana through one interface.

Mass distribution at scale

Up to 120 NFTs per transaction — a workaround for NEAR's throughput limit. A 10K-wallet campaign goes out in a few transactions, not hundreds. Automated airdrops and reward systems for collection launches.

Cross-chain by design

NEAR, TON, EVM and Solana collections work through one interface. Single tooling for cross-chain airdrops, drops and partner collaborations.

Marketplace, auctions & explorer

Sells, Orders and on-chain auctions inside the same service. A wallet explorer parses 10K+ addresses with full on-chain data extraction — holders, NFTs and tokens, all queryable.

Reputation game built in

Each NFT carries a positive or negative reputation coefficient. Wallets compete for TOP+ and TOP- rankings. Alchemy Lab — the cross-chain crafting mechanic — plugs in right here, turning distribution into a closed-loop economy.

#1 on NEAR by transaction volume. 1M+ tx on a single collection. Up to 120 NFTs per transaction. No analogs on the market.
Under the hood
  • Up to 120 NFTs per transaction — custom workaround for NEAR throughput limits
  • Cross-chain across NEAR + TON + EVM + Solana through one interface
  • Wallet explorer parses 10K+ addresses with full on-chain data extraction
  • Reputation engine — every NFT carries a positive or negative coefficient, TOP+ and TOP- leaderboards
  • Alchemy Lab cross-chain crafting plugs in directly — distribution becomes a closed-loop economy
Results
#1 on NEAR by tx volume 1M+ tx on a single collection 40 collections live 10K+ wallets in reputation NEAR · TON · EVM · Solana
Stack: TypeScript · NEAR SDK · Smart Contracts (Rust) · custom indexers · multi-chain interface (NEAR · TON · EVM · Solana)
AI · Pipeline

AI Content Factory

Python · GitHub Actions · ffmpeg · 10+ AI
Details →

AI Content Factory

AI · Pipeline
Role
AI engineering
Output
Video + slides from one prompt
Providers
10+ with fallback chain
Runtime
GitHub Actions CI

A production AI pipeline that turns a single prompt into finished video or presentation — scripts, voice, visuals, motion, music, effects, fully automated end-to-end. Fallback chains at every stage across 10+ AI providers, with a hybrid free/paid layer keeping cost predictable.

The challenge

Anyone can chain two AI APIs and ship a demo. Production content generation breaks the moment any single provider rate-limits, raises prices or quietly degrades quality. The pipeline needed to survive the failure of any single provider, balance free vs paid resources for cost-efficiency, and run unattended on CI — not in a tab.

What we built

A multi-stage pipeline (script → voice → visuals → motion → music → effects → assembly) where every stage has a graceful fallback to another provider. TypeScript + Node for orchestration, Python for AI calls and asset processing, ffmpeg + Pillow for media stitching. Runs on GitHub Actions as the execution engine — every generation is a CI job, fully reproducible. 10+ AI providers behind one router.

Fallback at every stage

Provider routing with automatic failover at script, voice, visuals, motion, music and effects layers — no single API can take down a generation.

CI-based execution

Runs as GitHub Actions jobs — unattended, reproducible, observable. Every generation is a CI run with logs, artifacts and retry semantics out of the box.

Hybrid free / paid routing

Free-tier providers wherever quality survives, paid providers where it doesn't. Cost stays predictable instead of bursting on every job.

Single prompt → finished media

Script + voice + visuals + motion + music + effects, chained automatically into a finished video or presentation. Prompt in, file out.

10+ AI providers, fallback at every stage, runs on CI — script → voice → visuals → motion → music → effects → assembly, all from a single prompt.

Under the hood
  • 10+ AI providers orchestrated through one router with per-stage automatic failover
  • Pipeline stages: script → voice → visuals → motion → music → effects → final assembly
  • Hybrid free / paid routing balances quality vs cost predictably
  • GitHub Actions as the execution engine — every run is a CI job, fully reproducible
  • ffmpeg + Pillow stitch generated assets into finished video and presentation outputs
  • TypeScript + Node for orchestration, Python for AI calls and asset processing
Results
10+ AI APIs orchestrated Fallback at every stage CI-based generation Single-prompt → finished video Hybrid free + paid routing
Stack: TypeScript · Node.js · Python · GitHub Actions · ffmpeg · Pillow · Gemini · DeepSeek · Claude · FLUX · LTX-Video · ElevenLabs
AI · No-code

YupJarvis — Ecosystem AI Agent Builder

Next.js · Supabase · AI SDK · React Flow
Details →

YupJarvis — Ecosystem AI Agent Builder

AI · No-code
Role
Full-stack product
Type
Studio product
Engine
Next.js 15 + Supabase
Scope
16 nodes · 7 agents · 302 assertions

YupJarvis is a no-code visual builder for LLM agents and pipelines — a drag-and-drop graph editor that chains 7 kinds of agents (LLM, Classifier, Extractor, Tool, Memory/RAG, Judge, Orchestrator), with triggers, public API, RAG over pgvector and a run journal with drill-down. Designed and built end-to-end by the studio.

The challenge

Building a production-grade LLM pipeline means stitching together agent types, retry logic, vector memory, tool calls, condition routing, run logs and access control by hand for every new project. The studio needed one substrate that turns this into a graph the team can draw, version and ship — not yet another framework wrapper.

What we built

A full Next.js 15 + Supabase application with a React Flow editor (16 node types and 7 agent kinds), JMESPath-powered Mapper/Condition routing, RAG on pgvector with an embedding pipeline, public API + webhook trigger, and a run journal with metrics, limits and drilldown UI. Configs validated through Zod, strict TypeScript across the stack, 302 test assertions across 11 test files — schemas, runner, classifier routing, tool calls, memory, judge, orchestrator and API.

Visual graph editor

React Flow with 16 node types, drag-and-drop wiring, Zod-validated configs per node — designers draw the pipeline instead of writing boilerplate.

7 agent kinds out of the box

LLM, Classifier, Extractor, Tool, Memory/RAG, Judge, Orchestrator. Each one a first-class node with typed schema, retries and observability.

RAG inside Supabase

pgvector storage, a separate embedding pipeline (OpenAI), retrieval as its own node type. Memory and RAG share the same substrate.

Run journal & public API

Every run logged with token and duration metrics, per-project limits, drilldown UI. Webhook trigger and public REST API expose any agent to the outside world.

16 node types, 7 agent kinds, 302 test assertions — built solo end-to-end on Next.js 15 + Supabase. Strict TypeScript, zero compile errors.

Under the hood
  • React Flow graph editor with 16 node types and Zod-validated per-node configs
  • 7 agent kinds: LLM, Classifier, Extractor, Tool, Memory/RAG, Judge, Orchestrator
  • JMESPath-powered Mapper and Condition nodes for routing and transformation
  • pgvector RAG inside Supabase with a separate embedding pipeline (OpenAI)
  • Webhook trigger + public REST API for invoking any agent from outside
  • Run journal — every step logged with token/duration metrics, limits and drilldown UI
  • 302 test assertions across 11 files: schemas, runner, classifier routing, tool, memory, judge, orchestrator, API keys, webhook
  • TypeScript strict, tsc --noEmit clean
Results
Built end-to-end solo 16 node types 7 agent kinds 302 test assertions 0 type errors
Stack: Next.js 15 (App Router · RSC · Server Actions) · Supabase (Postgres 16 + pgvector + RLS + Auth) · AI SDK 4.3 (Anthropic + OpenAI) · React Flow 11 · Zod · TypeScript strict
Web3 · Multi-chain

Alchemy Lab — Cross-Chain Crafting

NEAR · TON · React · bridge logic
Details →

Alchemy Lab — Cross-Chain Crafting

Web3 · Multi-chain
Role
Cross-chain game mechanic
Mechanic
Trade → craft → burn → mint
Chains
NEAR + TON
Recipes
317 live
cases/alchemy_title.jpeg

Alchemy Lab is a cross-chain NFT crafting mechanic unique on the market — players combine NFT components by recipes across NEAR and TON to mint brand-new NFTs in a closed-loop economy. Trade → craft → burn → mint, on two blockchains at once.

The challenge

NFT inflation kills value: once mass-minted, collections sit in wallets with nowhere to go. The fix had to make burning attractive — a game where players want to take NFTs out of circulation — and work across two blockchains so liquidity flows from both. And it had to be easy enough for partners to drop their NFTs into the same loop without rebuilding anything.

What we built

A custom recipe engine on top of NEAR + TON smart contracts and cross-chain bridge logic. Players combine NFT components by recipe to mint new NFTs — and the burned originals leave circulation forever. The same surface accepts partner NFTs, so external projects plug into the lab and get instant collab without writing their own crafting layer. 317 live recipes today, 34K+ crafts completed, TOP-2 on HOTCraft.

Cross-chain crafting

NEAR + TON in one closed loop. Liquidity from two blockchains feeds the same recipe library — combine components from either chain to mint new NFTs.

Burning solves inflation

Players themselves take NFTs and tokens out of circulation via crafting. Mass-minted supply is removed not by penalties but by a mechanic players choose.

Retention through recipes

317 recipes ship today — from basic to rare combos and boosts. Players come back daily to chase rare results.

Partner-ready surface

External projects drop their NFTs into the Lab — instant collab. The recipe engine is generic enough for any project to plug in without rebuilding crafting from scratch.

317 recipes, 34K+ crafts completed, two blockchains in one mechanic — the closed-loop economy that turns mass-minted NFTs into something players want to burn.
Under the hood
  • Cross-chain bridge logic uniting NEAR + TON liquidity into one recipe pool
  • Custom recipe engine — any NFT bundle definable as a craftable input
  • Burning + minting in one transaction per recipe
  • Partner-ready surface — external NFTs plug in without code
  • 317 recipes live, 34K+ crafts completed, TOP-2 on HOTCraft
Results
34K+ crafts completed 317 recipes live TOP-2 on HOTCraft NEAR + TON in one mechanic Partner-ready
Stack: NEAR/TON smart contracts · cross-chain bridge logic · React · custom recipe engine
AI · Vibe-coding

YupDar — AI Mini App for Self-Discovery

TG Mini App · Web · DeepSeek · Supabase
Details →

YupDar — AI Mini App for Self-Discovery

AI · Vibe-coding
cases/yupdar_title.jpeg
Role
Full-stack vibe-coding
Surfaces
TG Mini App + Web · RU / EN / ES
Team
Book author + Claude Code
Status
Live · trilingual · scientific track

A trilingual self-discovery ecosystem (RU / EN / ES) on the 64 Daram system — Telegram Mini App + a public web landing on one codebase, with a premium daily-companion layer ARKA on top. AI-generated literary portraits (8 psychological sections per user), oracle as a card spread, compatibility, hero's journey 2.0 with cross-user Daram unlocks, shadow quests, PvP card game. Books published in EN / ES on Amazon and on Litres in RU. A scientific track in parallel — DAR-27v2 research questionnaire, scientific-article DOCX builders, application to the World Science Championship in Dubai. Built solo by the book's author with Claude Code.

The challenge

Turning a published esoteric book on the 64 Darams into a living interactive product required AI capable of writing literary Russian indistinguishable from the author's voice. Off-the-shelf models produced "professional" Russian — methodical, cliché-laden, sprinkled with Latin tokens, with broken gender agreement. Declarative bans inside the prompt didn't hold. And the product had to ship as both a Telegram Mini App AND a public SEO-landing — one codebase, two surfaces.

What we built

A two-layer Russian-AI quality stack: prompt v7 with the author's ~600-word reference text baked in as a stylistic anchor, plus a programmatic postprocessor russifyText() with 60+ regex defenses against forbidden constructions (clichés, Latin, broken gender, formulaic openings). Provider routing — DeepSeek-V3 primary (~$0.0024 per message), Groq Llama 3.3 70B as automatic fallback, Llama 3.1 8B as second fallback. response_format: json_object guarantees parseable output for 8-section portraits. Yandex Speller post-pass with a protected glossary. Same backend serves both the Mini App and the yupdar.com landing — split only at the HTML entrypoint level. On top of the base game the studio shipped a premium ARKA layer (a daily companion in the rhythm of the user's Daram — power windows, archetype-voiced rituals, AI strategy assistant, mentor brief, content plan), Hero's Journey 2.0 (a viral mechanic to unlock the Darams of close ones via referrals, crystals or direct payment), a Pricing v2 subscription system (1/3/6/12 months, up to −33% off, −50% first-month promo, YuKassa-wired), and a secondary doTERRA affiliate channel (essential oils mapped to each Daram). Full localization rolled out to English and Spanish across the entire interface, the encyclopedia and the in-app book reader (RU 94 / EN 99 / ES 87 chapters). In parallel a scientific track is running — the DAR-27v2 research questionnaire, DOCX builders for scientific articles, an application submitted to the World Science Championship in Dubai.

Literary AI portraits, trilingual

8-section JSON output per user — Archetype, Essence, Superpowers, Flow, Relationships, Traps, Key to Self, Purpose. ~5K–6K chars of structured literary text per generation. Fully localized to Russian, English and Spanish across portraits, oracle and encyclopedia.

ARKA premium daily layer

A daily companion in the rhythm of the user's Daram, sitting above the base game. Power windows, archetype-voiced rituals tied to doTERRA essential oils, task tracker with load types, AI strategy assistant, mentor brief, finances, content plan, week review — all in the archetype's voice. Prototype isolated in a sandbox so prod stays untouched.

Hero's Journey 2.0 — viral mechanic

Users now unlock the Darams of close ones (partner, friends, parent of a child) via referrals, crystals or direct payment. Each user becomes a conductor into YupDar for the people around them — and a new revenue surface.

Scientific track in parallel

DAR-27v2 research questionnaire, DOCX builders for scientific articles, application submitted to the World Science Championship in Dubai. A dedicated "Science 🔬" tab in the new architecture preview. International recognition of the Daram system is an explicit goal.

Trilingual (RU / EN / ES). Premium ARKA layer on top of the base game. Viral cross-user Daram unlocks. Pricing v2 subscriptions. doTERRA affiliate. Scientific track to the World Science Championship in Dubai. Books on Amazon and Litres. All shipped solo by a book author through Claude Code.

Under the hood
  • DeepSeek-V3 primary, Groq Llama 3.3 70B fallback, Llama 3.1 8B as second fallback
  • response_format: json_object — guaranteed valid JSON for 8-section literary portraits
  • russifyText() programmatic postprocessor — 60+ regex defenses against forbidden constructions
  • Author's ~600-word reference text baked into prompt v7 as a stylistic anchor
  • Yandex Speller post-pass with a protected glossary for system terms
  • Full RU / EN / ES localization across interface, oracle, encyclopedia and the in-app book reader (RU 94 / EN 99 / ES 87 chapters)
  • ARKA premium layer: a daily companion in the rhythm of the user's Daram — sandboxed prototype isolated from prod (iram-prototype/)
  • Hero's Journey 2.0 — viral mechanic to unlock close-ones' Darams via referrals / crystals / direct payment
  • Pricing v2 — 1/3/6/12-month subscriptions with discounts up to −33%, −50% first-month promo, YuKassa across all tiers and periods
  • doTERRA affiliate channel — essential oils mapped to each Daram, second monetization line
  • DAR-27v2 research questionnaire, DOCX builders for scientific articles, application submitted to the World Science Championship in Dubai
  • 11 Vercel serverless endpoints; content.js routes 6 generation types
  • Supabase Postgres — 11+ tables, 11 SQL migrations, RLS + service-role split
  • HMAC-SHA256 validation of Telegram initData; anonymous web sessions with migration into Telegram
  • PDF.js-powered in-app reader for the 212-page authored book (RU / EN / ES)
  • Telegram Stars admin tab with live bot balance via API + DB reconciliation
  • Architecture preview rebuild — new 4-tab nav (Я / Семья / Сокровищница / ARKA) + Science 🔬 header, deployed to a separate preview surface
  • 41 "project laws" in Claude Code memory system
Results
Trilingual · RU / EN / ES Premium ARKA layer live Hero's Journey 2.0 viral mechanic Pricing v2 subscriptions doTERRA affiliate DAR-27v2 + Dubai application Books on Amazon + Litres Built solo by book author
Stack: Vanilla JS · HTML5 · Telegram Mini App SDK · PDF.js · Node.js · Vercel Serverless · Supabase (Postgres) · DeepSeek-V3 · Groq Llama 3.3 70B / 3.1 8B · Yandex Speller · i18n (RU / EN / ES) · YuKassa subscriptions · doTERRA affiliate · Telegram Stars API · DOCX builders for scientific articles · GitHub worktrees → Vercel preview deployments
Automation

Yupland Lottery — Live Stream Raffle

TypeScript · Node.js · Telegram Bot API
Details →

Yupland Lottery — Live Stream Raffle

Automation
Role
Automation engineering
Use case
Live-stream raffles
Surface
Telegram + admin dashboard
Replaces
Manual stream-side ops

An automated service for live-stream raffles — replaces the manual stream operator with a one-click pipeline. Streamer picks parameters, the system runs a fair randomizer, delivers the result through Telegram, syncs to a real-time admin view and renders output in the streamer's custom format.

The challenge

Running a live raffle on a stream is fragile manual work — the operator collects entries off-camera, randomizes them in a spreadsheet, then tries to present the result believably. It eats time during the stream, breaks the moment anyone questions the randomness, and doesn't scale across multiple streamers running the same flow.

What we built

A one-click automation: a TypeScript + Node service collects entries through a Telegram bot, runs a fair randomizer with verifiable output, syncs the result in real time to an admin dashboard the streamer reads from, posts the winner back to Telegram and renders results in a custom on-camera format. The streamer goes from "okay let me pull up the spreadsheet" to a single button press.

Fair randomizer

Verifiable output that takes the trust question off the table during a live stream — no "wait, who picked the winner?" debates in chat.

Real-time event sync

Admin dashboard updates as entries arrive; the streamer reads from one view instead of switching tabs mid-stream.

Telegram-native delivery

Entries in, winner out, all through a Telegram bot. No external accounts to manage, no extra channels to publish to.

Custom result formatting

Output renders in the streamer's own style — ready to read on-camera without copy-paste gymnastics.

One click on stream, one fair winner — replaces what used to be a manual back-stage scramble.

Under the hood
  • Fair randomizer with verifiable output
  • Telegram Bot API integration for entries and result delivery
  • Admin dashboard with real-time event sync
  • Custom result formatting per streamer
  • One-click automation replaces manual stream-side raffle ops
Results
One-click automation Real-time event sync Telegram delivery Fair randomizer Custom result formatting
Stack: TypeScript · Node.js · Telegram Bot API · real-time event sync
Web3 · Game

Duplo — On-chain Combo Game

TypeScript · NEAR + TON · parser engine
Details →

Duplo — On-chain Combo Game

Web3 · Game
Role
Web3 game + on-chain parser
Mechanic
Burn → combo → reward
Versions
Manual + automated
Chains
NEAR + TON
cases/duplo_title.jpeg

Duplo is the burn-and-reward game at the heart of Yupland — players send NFTs and tokens to the Duplo wallet, the engine matches them by recipe and pays out rewards on-chain. Two implementations: the original game (manual play, with a rating) and the version integrated into YupLand Portal (automated, no rating, runs inside the daily loot loop).

The challenge

An ecosystem with mass-minted NFTs needs a sink — a way to take supply out of circulation that feels like a game, not a penalty. Players also need a reason to actively combine the NFTs they own. The same mechanic had to work two ways: as an original game (manual play, leaderboard) and as an automated daily loop inside a larger portal (no rating, broad reach).

What we built

A burn-and-reward engine with two faces, sharing one on-chain wallet and one recipe library. Original game: a custom blockchain parser watches transfers to the Duplo wallet in real time, identifies multi-asset combinations by recipe (NFT + token bundles, time-window logic), computes rewards and maintains a leaderboard. Portal version: the same combo logic but fully automated inside YupLand Portal — players don't send transfers manually; the portal triggers burn → combo check → reward inside its daily loot loop. Both versions feed the same wallet, putting supply out of circulation across the whole ecosystem.

Original game (with rating)

Players send NFTs and tokens to the Duplo wallet manually. A custom parser identifies valid combinations in real time, computes rewards and maintains a leaderboard.

Portal version (automated)

Same combo logic, no manual transfers. YupLand Portal runs burn → combo check → reward inside its daily loot loop. No rating — designed for breadth, not competition.

Multi-asset combo engine

Detects bundles of NFTs + tokens by recipe. Time-window logic, configurable reward tables, multi-chain (NEAR + TON).

Shared on-chain sink

Both versions feed the same Duplo wallet — a fun supply sink at scale. Players take NFTs and tokens out of circulation because they want to, not because they have to.

Two implementations, one wallet, one recipe library — the burn-and-reward loop that sucks NFT supply out of circulation across the whole ecosystem.
Under the hood
  • Real-time on-chain parser watching the Duplo wallet for incoming bundles
  • Multi-asset combination detection (NFT + tokens) with time-window logic
  • Shared recipe library across both implementations
  • Automated burn → combo → reward path inside YupLand Portal (no manual tx)
  • Configurable reward tables, multi-chain (NEAR + TON)
Results
~2M boxes / 30d (portal version) Real-time on-chain parsing Multi-asset NFT + token combos Shared sink · 2 implementations Multi-chain (NEAR + TON)
Stack: TypeScript · NEAR + TON RPC · custom parsing engine · recipe library · YupLand Portal integration
Web3 · Pipeline

NFT Collection Launch Service

NEAR · TON · image processing · Node.js
Details →

NFT Collection Launch Service

Web3 · Pipeline
cases/nft-launch_title.jpeg
Role
Web3 infrastructure
Type
Reusable studio pipeline
Chains
NEAR + TON
Status
Used in multiple launches

A full pipeline for launching custom NFT collections on NEAR and TON — from art assembly to on-chain deploy. Trait generator from art components, smart contract templates, custom token tooling and a mint flow integrated with launch landings. Reused across multiple collection launches by the studio.

The challenge

Every NFT collection launch hits the same pipeline: assemble traits into final images, deploy a smart contract, build a mint UI, wire it into a landing. Doing it from scratch each time burns weeks per launch. The studio needed a reusable pipeline that takes art components in and delivers a live mint on the other end.

What we built

An end-to-end pipeline: a trait generator stitches art components into final NFT images by rarity rules, a smart contract template handles NEAR or TON deployment, custom token tooling manages token-side configuration, and the mint flow is integrated with launch landings as a drop-in component. Art in → button → live mint out.

Trait generator

Combines art components into final NFT images by rarity rules. Same input layer feeds both chains.

Multi-chain smart contracts

NEAR and TON contract templates plugged into the same pipeline — one configuration, two deploys.

Custom token tooling

Token-side configuration managed alongside contract deploy — collection metadata, supply, royalties.

Mint flow + landing

Drop-in mint component for the launch landing — wallet connect, mint button, supply counter, all wired in.

Art → contract → mint, multi-chain, reused across multiple launches by the studio.

Under the hood
  • Trait generator stitches art components into final NFT images per rarity rules
  • Smart contract templates for NEAR and TON, plugged into the same pipeline
  • Custom token tooling for token-side configuration (metadata, supply, royalties)
  • Mint flow integrated as a drop-in component for launch landings
  • Pipeline reused across multiple collection launches
Results
Used in multiple launches NEAR + TON Art → contract → mint Drop-in mint component
Stack: TypeScript · NEAR/TON smart contracts · image processing · Node.js
AI · Telegram

AI Image Generation Bot

Telegram · Supabase · image gen · crypto
Details →

AI Image Generation Bot

AI · Telegram
Role
AI engineering
Platform
Telegram bot
Type
Client product
Timeline
10 days

A production Telegram bot that generates stylized personal images while keeping the user's identity recognizable across every generation.

The challenge

Preserve a recognizable face through heavy stylization, keep generation costs under control and handle bursty async load — all shipped fast.

What we built

A multimodal input layer, dual-API failover for cost optimization, an async job architecture and crypto payments — live in 10 days.

Results
10 days to production Dual-API failover Crypto payments
Stack: Node.js · Telegram Bot SDK · Express · Supabase · Image gen + fallback · STT · Blockchain payments
// Technology

Full-stack across
AI, Web3 and product

We pick the right tool for each problem and integrate up to 10+ providers in a single pipeline with graceful fallback.

LLM Providers

Claude (Anthropic) GPT (OpenAI) DeepSeek-V3 Gemini 2.5 Groq Llama 3.3 70B Llama 3.1 8B

Generative APIs

Suno (music) FLUX Replicate Kling Pollinations.ai LTX-Video Luma Dream Machine Fal.ai ElevenLabs

Voice & Audio

Microsoft edge-tts ElevenLabs Whisper faster-whisper Word-level timestamps

Embeddings & RAG

Voyage AI pgvector OpenAI embeddings Hybrid search

Frontend

React Next.js 15 React Flow Vite TypeScript Tailwind shadcn/ui Vanilla JS PDF.js

Backend

Node.js Express Python Vercel Serverless AI SDK 4.3 Zod Render.com GitHub Actions Sentry

Database

PostgreSQL Supabase pgvector RLS & Auth sql.js (local)

Web3 / Blockchain

NEAR Protocol TON Ethereum / EVM Solana Rust contracts Custom indexers Smart contract templates

Telegram

grammY Bot API (raw) Mini App SDK HMAC-SHA256 auth

Payments

YuKassa (RU acquiring) Crypto payments GetCourse API Card acquiring (regulated)

Mobile

Capacitor 8 iOS + Android single codebase Native bridges Push notifications

Tools & Infra

ffmpeg Pillow / PIL Worktree dev flow Vercel Preview Claude Code memory system
// The team

Seven people.
Each has shipped their own product.

Not a freelance marketplace, not a generic dev shop. Everyone on the team has built and runs a product in our own Yupland ecosystem — so client work gets the same level of ownership as our own.

A Angelina

Angelina

// Founder · CCO

Founder of YupCode and Yupland, concept designer of the ecosystem, public face of the team.

A Alla

Alla

// AI Engineer

Built AI Music Platform and AI Image Bot. Co-author of YupSelf. Multimodal pipelines and Mini Apps.

M Mikhail

Mikhail

// Web3 Developer

Telegram bots, NFT infrastructure, P2E games, blockchain parsing. AI Content Factory creator.

Y Yaroslav

Yaroslav

// Lead Game Designer

Game design and economy across the ecosystem. Co-author of YupSelf AI avatars.

I Ilya

Ilya

// Full-stack Engineer

Creator of YupLand Portal with DEX, YupPay, YupXBet, YupJarvis and other services.

R Roman

Roman

// Web3 Backend

NEAR indexers, smart contracts, NFT Sender (#1 distribution on NEAR), marketplace infrastructure.

S Svetlana

Svetlana

// Vibe-coder · Author

Built YupDar as a non-developer book author via Claude Code. Living proof that vibe-coding ships products.

+

And a network

// Specialists on call

Designers, translators and domain experts we bring in when project scope requires it.

// FAQ

Frequently asked

The things B2B partners and corporate clients ask before saying yes. Click a question to expand.

Will it survive production load? Can vibe-coders handle that?

Speed comes from vibe-coding, but underneath sits classical full-stack engineering. We understand databases, load behavior, security and the code itself — not just how to prompt Claude.

  • Two-environment CI/CD — staging and production, healthchecks, rollback playbooks, documented deploy.
  • Schema design and indexing from day one — no "works on 10 users, dies on 1,000" surprises.
  • Load patterns we've already survived — Yupland ecosystem with 5.7M+ on-chain transactions, NFT-Sender at #1 on NEAR by volume, YupLand Portal with ~2M loot boxes / 30 days.
  • Real engineers behind the AI flow — the team has years of production work before AI accelerated us. Vibe-coding is the multiplier; engineering is the base.
Can we start with a small pilot before committing?

Yes — that's the recommended way to start.

  • Discovery call is free. 30 minutes, no sales pitch, no obligation. You leave with a realistic scope, timeline and price.
  • MVP scope is welcome. We can take a single narrow piece (one bot flow, one landing, one AI agent) and ship it in 1–2 weeks before any larger commitment.
  • Hourly with weekly cap works if the scope is still moving. You see exactly what time was spent on what.
  • Milestone-based payment for fixed-scope projects. You only pay for what's delivered.
Who owns the code after the project?

You do — fully. We don't try to lock you in.

  • Full repository transfer on completion — your GitHub/GitLab, your account.
  • Documented handover — env notes, secrets list, deployment instructions, rollback playbook, 30-minute walkthrough call.
  • No proprietary "framework" that ties you to us — we build on standard, popular stacks any team can pick up.
  • Your team or another contractor can continue the work without us. We'd rather you come back because you want to, not because you have to.
What if something breaks after handover?

Production discipline is baked in from the start — we plan for things breaking so you don't scramble when they do.

  • Rollback playbook ships with the project — one command brings the previous version back.
  • Healthchecks and alerts set up before launch — you know about an outage before users do.
  • Two-environment CI/CD — every change tested on staging before reaching production.
  • Post-launch support available on a separate agreement — hourly retainer or fixed monthly. We help during incidents even when there's no retainer; we just bill the hours.
What if the AI makes a mistake in production?

AI gets things wrong. The architecture is what determines whether those mistakes hurt you.

  • Human approval on irreversible actions — payments, deletions, outbound communication, prod config changes always need a person to click "approve". Autonomy is reserved for read-only and reversible operations.
  • Provider fallback chain — if one AI provider returns nonsense or goes down, we automatically retry on another. No silent failures.
  • Validation on every tool call — schema checks, business-rule checks, argument whitelists. The model can't "ask" to do something unsupported.
  • Full audit log — every AI decision is traceable. If something does go wrong, we know exactly what the model saw, what it decided, and why.
How do you keep our data safe?

For mid-market and corporate clients in regulated markets (152-ФЗ, GDPR, PII, financial data) security is the #1 criterion. We treat the model as an untrusted user — rights live in a deterministic tool layer with audit and human approval on irreversible actions.

  • LLM ≠ a service with rights. Every action goes through a deterministic tool layer that checks permissions in the context of a specific user. A prompt injection can't "ask" to delete data — the policy engine rejects the call.
  • Sensitive data → internal contour. Local model (vLLM / Ollama on your GPUs) or enterprise tiers with Zero Data Retention. Public LLMs never see raw data. Sensitivity-aware routing through an AI Gateway.
  • Audit on every action. Each tool call written to an immutable log: who, what, when, which prompt, what result, allowed or denied. Alerts on anomalous patterns.
  • Human approval on irreversible. Payments, data deletion, outbound mail, prod config changes — always through an approval queue.
What we deliberately don't do:
  • × Give the model raw run_sql(query)
  • × Run agents under an admin service account
  • × Defend ourselves with a system prompt only
  • × RAG without document-level ACL
  • × Log raw prompts containing PII or secrets
  • × Direct model → DB / CRM / ERP connection
Working knowledge of: 152-ФЗ (RF) GDPR ISO 27001 controls OWASP LLM Top 10 SOC 2 patterns

Full engineering brief (40-page guide) available on request.

// How we work

From inquiry to handover
in four steps

Predictable process, predictable outputs. No surprises, no scope creep without a written change.

01 — DISCOVERY

30-minute call · free

We read your brief, ask sharp questions, push back on weak assumptions. You leave with a realistic scope, timeline and price — no sales pitch, no obligation.

02 — SCOPE

Written agreement

Fixed deliverables, fixed timeline. Or hourly with a weekly cap if the scope is still moving. Payment by milestone — you only pay for what's done.

03 — BUILD

Daily updates

Progress every day in your preferred channel (Telegram, Slack, email). No "find out two days before deadline that nothing is ready."

04 — HANDOVER

Documented delivery

Deploy-ready code, env notes, rollback playbook, 30-minute walkthrough call. Your team keeps building or takes over without us — we don't try to lock you in.

Have a defined scope but no engineering bandwidth?

Send a 2–3 sentence brief and we'll come back with a realistic scope, timeline and price within 24 hours. Discovery call is free.

Book a 30-min discovery call →