· 14 мин чтения
AI гораздо больший человек, чем вам кажется
Забывает, спешит, читает правила по диагонали, теряет фокус под нагрузкой. Почему идеальная работа с AI начинается с права на ошибку — и что об этом говорит наука.
> 10 лет руководил продуктами. Теперь строю их с AI. 10 лет управлял продуктами и командами. Теперь строю AI-продукты.
> 39 продуктов в проде. Выручка ×2.2. Команда из 150 → solo + 15 AI-агентов. 39 продуктов в продакшене. Выручка ×2.2 за год. 150 человек → один + AI.
> Ниже — каждый продукт, каждая подсистема, каждая технология. Ниже — портфолио, технологии, статьи. Всё можно проверить.
Руководил 150 людьми, 11 командами, P&L на сотни миллионов. Вырастил выручку ×2.2 при падении рынка на 37%. Потом понял: один человек с правильной AI-архитектурой делает больше, чем отдел.
Построил Factory OS — систему из 15 AI-агентов (Builder, Quality, DevOps, Discovery, Strategy...), которые пишут код, тестируют, деплоят и чинят баги. Я принимаю решения. Они исполняют.
Результат: 39 продуктов в продакшене. Флагман — AICPO (aicpo.ru): платформа из 31 подсистемы для продуктового исследования с AI. Граф знаний, мультимодельный LLM-каскад, self-evolving промпты, competitive intelligence.
Всё что ниже — доказательство. Не резюме, не обещания. Рабочий код, живые продукты, проверяемые результаты. Product Lead в PUNKT E — крупнейшее EV-приложение в России, 95 000 пользователей. Вырастил выручку в 2.2 раза за год при падении рынка на 37%.
Head of Products в Mafin — Top-3 insurtech: 11 команд, 150 человек, P&L четырёх направлений, расширение линейки с 3 до 16 продуктов.
Сейчас: строю AI-продукты полного цикла — от исследования рынка до работающего сервиса. 39 продуктов в продакшене. Флагман — AICPO: платформа из 31 подсистемы, которая помогает продуктовым командам принимать решения на данных.
Открыт к руководству AI-продуктами, консалтингу и корпоративному внедрению AI.
## Indie Products ## Самостоятельные продукты
Astro, Next.js, Python, Cloudflare — standalone Отдельные продукты, запущенные в продакшен
CoffeeBot
LIVE
AI coffee quiz AI picks coffee based on taste preferences
Astro Groq Tailwind 100+ profiles quiz
82 coffee profiles generated 82 recommendations, 100+ varieties
InkCloak
LAUNCHING
AI text detector + humanizer Detects AI text with 99.5% accuracy
Next.js DeBERTa LoRA fine-tune RunPod GPU AUROC 0.9948 CUDA
AUROC 0.9948 detection accuracy 99.5% accuracy, GPT/Claude/Gemini
Scrutly
PLANNED
Free AI text detector Free AI text check, no registration
Astro DeBERTa free
Tool Sites
LIVE
199 SEO pages across 5 domains 199 niche tools for SEO traffic
Astro MDX programmatic SEO 5 domains 199 pages
199 pages across 5 domains 199 pages, 5 sites
OpenClaw Finance
BUILT
AI agent factory for financial analytics AI agents for financial analytics and consolidation
Python DeepSeek R1 Docker 5 agents multi-agent
5 agents, deployed on 2 machines 5 agents, automated reports
RackCloud
BUILT
Cloud platform architecture for 400+ servers Cloud platform project for 400+ servers
Claude Opus Proxmox 400+ servers 17 services 10 racks
17 IaaS services, 400+ servers, 10 racks 400+ servers, 17 services
DIYCalc
LIVE
Construction material calculators Construction material calculators
Astro MDX programmatic SEO 30+ calculators
30+ calculators 30+ calculators
FitCalc
LIVE
Health & fitness calculators Health & fitness calculators
Astro MDX programmatic SEO 40+ calculators
40+ calculators 40+ calculators
## Corporate / AI Engine ## Corporate / AI-движок
LLM cascade, vector search, knowledge graph, autonomous research — AICPO build Модели, память, поиск по смыслу, исследование рынка — сборка AICPO
Autonomous LLM Runtime
LIVE
7 fallback levels, auto model calibration, 0 downtime Smart routing between AI models: auto-selects best by speed and cost
Groq OpenRouter Claude Gemini Qwen Llama Mixtral 7 fallback levels auto-blacklist
7 fallback levels, 0 downtime auto model selection, budget control
Dual-Brain Chat
LIVE
Fast brain for simple, smart for complex. Auto-switches Fast brain for simple, smart for complex. Auto-switches
Groq 0.5s Claude load balancing escalation de-escalation -90% tokens
5 triggers, 0.5s → Claude fast + smart, -90% cost
Semantic Memory
LIVE
Vector search by meaning + auto-tracking 19 AJTBD data points Platform remembers everything: search by meaning, not keywords
pg_vector cosine similarity 19 data points project brief
19 data points, vector search semantic search, data tracking
Vector Intelligence Engine
LIVE
4-level data model: text → facts → vectors → graph Multi-level analysis: from raw text to knowledge graph automatically
text→facts→vectors→graph pg_vector 5 dimensions SQLite fallback
4 levels, pg_vector, 5 dimensions 4 levels of data processing
Knowledge Graph
LIVE
Graph storage: 14 node types, evidence trail to user words Knowledge graph: connections between facts, trace to client words
14 node types 18 edge types confidence scoring evidence trail Rails engine
14 node types, edges, evidence trail 14 object types, traceability
AI Discovery Engine
LIVE
Autonomous research: search → TG channels → articles → LLM filtering Automatic market research: channels, articles, insights by niche
DuckDuckGo Telegram vc.ru Habr 3 rounds cache 90d
3 rounds, cache 90d automatic research in minutes
## Corporate / Product Intelligence ## Corporate / Продуктовая аналитика
Artifacts, fact extraction, competitors, trends — AICPO build 43 документа, мониторинг конкурентов и трендов — сборка AICPO
AICPO Platform
LIVE
AI workspace for product research. 31 subsystems, 43 artifacts AI assistant for product research. 31 subsystems, 43 documents
Rails 8 PostgreSQL Docker ActionCable SSE 31 subsystems 43 artifacts
43 artifacts, 31 subsystems 43 documents, 31 subsystems
Artifact Pipeline
LIVE
3-stage generation: Structure → Generate → Validate → retry 43 strategic documents: from JTBD maps to financial models
Groq→Gemini→Groq structure generate validate retry 43 types
43 types, 3 stages 43 documents with quality check
Pain Extractor
LIVE
Chat export parser → 19 AJTBD data points Upload chat export → get structured analysis
Rails engine Groq TG Discord Slack WhatsApp 19 fields
4 messengers → 19 fields 4 messengers, auto-analysis
Research Corpus
LIVE
Shared knowledge base: one research — data for everyone Shared research base: research once — data available for all
PostgreSQL SHA256 dedup cross-user TTL 90d
cross-user, TTL 90d shared base, no repeated searches
Competitive Intelligence
LIVE
Competitor monitoring: auto-scrape, signals, digest Competitor monitoring: automatic tracking of changes and threats
DuckDuckGo web scrape Groq KG integration cron 6h weekly digest
auto-scrape 6h, KG integration automatic competitor monitoring
Trend Monitoring
LIVE
Trend monitoring: DuckDuckGo → clustering → KG → 3 artifacts Trend monitoring in niche: automatic search and impact assessment
DuckDuckGo Jaccard similarity Groq KG weekly 3 artifacts
weekly scan, 3 artifacts weekly trend search
## Corporate / Self-Evolution ## Corporate / Самообучение
Auto-improving prompts, feedback loops, incidents → features — reusable modules Система учится на ошибках и развивается без участия человека
Self-Evolving Prompts
LIVE
Prompts improve themselves from negative feedback System learns from mistakes: AI auto-improves answer quality
weekly cron Groq auto-draft versioning rollback A/B prompts
auto-draft weekly self-learning from reviews
Evolution Pipeline
LIVE
AI generates backlog from feedback and chat audit Automatic roadmap: AI analyzes feedback and prioritizes improvements
Claude Sonnet Groq daily digest weekly analysis voting prioritization
Feedback → AI → Backlog AI-prioritized improvements
Ship Pipeline
LIVE
Vote → promote → auto-generate → notify. Product evolves itself Users vote for features → AI generates → product evolves itself
ActionCable Groq vote ≥3 auto-generate notify
vote → promote → auto-generate voting → auto-generation
Quality Management
LIVE
Quality management: feedback → improvement → versioning Quality management: reviews → improvement → versioning
accuracy completeness relevance staleness detection prompt versioning
3 dimension feedback, auto-draft 3 parameters, auto-improvement
Incident → Feature
LIVE
Incident escalation → features. AI-managed backlog Bug → analysis → feature: AI turns incidents into improvements
Claude Sonnet Groq 8 categories chat audit auto-backlog
incident → AI analysis → feature bug → AI analysis → improvement
## Corporate / Enterprise ## Corporate / Корпоративные функции
RBAC, SSO, CRM, billing, analytics, admin — reusable modules Командная работа, тарификация, CRM, маркетинговая аналитика
Team OS
LIVE
Organizations, RBAC, SSO (SAML + OIDC), audit log Teamwork: roles, SSO authentication, action audit
SAML 2.0 OIDC RBAC 4 roles auto-provisioning audit log @AI comments
SAML + OIDC, 4 roles roles, SSO, audit
CRM Engine
LIVE
CRM: lead scoring, churn detector, reactivation, referral CRM: lead scoring, churn detection, reactivation, referral program
AI lead scoring churn detector reactivation referral 28 campaigns
lead scoring, churn, referral scoring, churn, referral program
Credit Engine
LIVE
Atomic transactions with row lock — race condition impossible Billing: 6 tiers, transparent spending, overdraft protection
6 tiers atomic deductions row locks BYOK soft limits overdraft
6 tiers, row locks 6 tiers, transparent spending
Tokenomics Guard
LIVE
Per-request LLM cost tracking, budget gate AI spending control: daily limit, alert when approaching ceiling
per-request tracking $5/day limit alert 80% auto-block premium
$5/day, alert 80% AI budget control
P&L Economics
LIVE
P&L and unit economics of the platform P&L and unit economics of the platform
6 tiers per-member analytics transaction history usage %
6 tiers, soft limits, P&L 6 tiers, spending analytics
User Management
LIVE
User management: invites, tiers, anonymous funnel User management: invites, tiers, registration
invites 3 tiers shadow projects claim IP rate limit welcome bonus
invites, 3 tiers, shadow projects invites, tiers, registration
Admin Command Center
LIVE
Command center: KPI, CRM, analytics, Ship pipeline, TG channel, invites Command center: KPI, CRM, analytics — all in one
KPI bulk actions hotkeys CRM backlog TG publish
KPI, bulk actions, hotkeys unified management dashboard
Marketing Analytics
LIVE
Funnel, cohorts, campaigns, UTM tracking Funnel, cohorts, campaigns, UTM tracking
PostHog GA4 UTM funnel cohorts 28 campaigns weekly reports
funnel, cohorts, 28 campaigns funnel, cohorts, campaigns
## Corporate / Integrations ## Corporate / Интеграции
Monitoring, email, PDF, TG, blog, health checks — reusable modules Мониторинг метрик, email, документы, публикации
Product Health Monitor
LIVE
Live product pulse: 46 integrations, 6 live connectors, 28 in roadmap Business metrics monitoring: PostHog, Stripe, GA4, Metrika — in one dashboard
PostHog Stripe GA4 Metrika YooKassa CloudPayments 46 integrations
46 integrations, 28 connectors 46 integrations, live dashboard
Email Bridge
LIVE
Bidirectional email channel: IMAP poll → project routing Email integration: send and receive emails from the platform
IMAP poll 2min SMTP ActionCable auto-routing bounce protection
poll 2min, auto-routing incoming + outgoing, auto-routing
PDF Publisher
LIVE
Public PDFs with digital signature: document ID + frozen content + verification Document publishing: PDF with signature, authenticity verification, public links
document ID content freeze verification B4 print-ready watermark
B4, watermark, public links PDF with signature, verification
TG Channel Publisher
LIVE
Content publishing to TG channel from platform Telegram channel publishing from platform
CF Worker relay Telegram API preview scheduled posts
CF Worker relay, TG channel publishing from admin
Blog Publisher
LIVE
Blog system with publishing from platform Blog system with publishing from platform
Markdown SEO preview admin panel
markdown, SEO, preview articles, SEO, preview
System Health
LIVE
System health: health check, deploy status, monitoring System health monitoring: checks, status, alerts
health check deploy status alerts diagnostics weekly cron
health check, deploy status monitoring, alerts, diagnostics
## Последнее
· 12 мин чтения
AI идёт в разнос: почему чем дольше разговор, тем хуже результат
Ошибочные цепочки рассуждений получают вес в контексте и начинают фонить на все последующие решения. Почему AI блестяще решает с нуля, но не может выбраться из ямы. Научное объяснение + практика.
· 11 мин чтения
Почему AI-агенты ошибаются в финансах и как я это починил
AI — не калькулятор. 90% точность одного агента × 90% второго = 81% на выходе. В финансовой консолидации это катастрофа. Как я компенсировал вероятностную природу LLM циклами проверки.
## Рассылка
Следите за проектами
Одно письмо когда выходит новый продукт. Без спама.