AI asosiy qism, o'ram emas

Pitch «biz AI ishlatamiz» emas. Dunyoning yarmi AI ishlatadi. Pitch shundaki, AI bizning stack'imizda paydo bo'lgan har bir joyda, AI funksiyani ishga tushiradigan narsadir, AI'ni olib tashlasangiz, funksiya shunchaki yomonlashib qolmay, balki butunlay ishdan chiqadi. Bu chatbot-on-top-of-product'dan boshqa kategoriya va uni nusxalash qiyin.

AI qayerda asosiy (imkon beruvchi emas)

Designer Studio

• Bitta orchestrator ostida 5 ta tashqi AI providers (Anthropic, OpenAI, recraft, replicate, quiver) bo'ylab 14 ta atomic tools.
• Claude Sonnet ko'p bosqichli pipeline'ni ishga tushiradi: analyze → extract → vectorize → recurse → compose.
• Multimodal: agent oraliq natijalarni KO'RADI va keyinroq recurse qilishni, muqobil yo'llarga qaytishni yoki foydalanuvchidan so'rashni hal qiladi.
• Agent darajasida qaror qabul qilish: «bu trace yetarlicha toza bormi?» → recurse / edit_image fallback / human prompt.
• «AI rasm yaratadi» emas, AI 14-vositali brand-identity pipeline'ni boshqaradi.

Market Research AI

• GPT-4o fallback bilan Claude Sonnet 4.
• 121 ta production tables va 60+ event'li PostHog event-stream'ga ega jonli PostgreSQL database ustida Tool-use.
• Agent SQL'ni tezda yaratadi, uni ishga tushiradi, natijalarni tasdiqlaydi, boy javob bloklarini sintez qiladi.
• Har bir surface uchun ko'rsatilgan boy response blocks bilan Streaming SSE (web → full blocks; Telegram → ikkinchi parser-prompt orqali TG-friendly).
• «Hech qachon raqamlarni uydirma» system prompt'da majburlanadi.
• Bu chat-with-our-docs emas. Bu haqiqiy production schema'da jonli data ishini bajaradigan agent.

Social Tracker

• 3-tier cost-aware pipeline: 1. Claude Haiku, caption screen (cheap) 2. Claude Haiku-vision, thumbnail check (cheap-ish) 3. Claude Sonnet + Whisper, deep video and audio analysis (expensive, gated)
• Hamma narsani eng og'ir tier orqali ishlatishga nisbatan taxminan 10× cost reduction.
• Har bir app uchun Custom product-context system, xuddi shu pipeline fitness app uchun dating app'dan farqli ishlaydi.

Builder API

• AsyncAnthropic, 5 analyzer endpoints.
• creators'ning GitHub repos'ini o'qiydi → SDK events / funnel / push / referral hooks'ni ajratib oladi.
• Multi-repo analysis uchun Parallel-gather.
• Bu cross-ecosystem porting'ni amalga oshiradigan qatlam, same repo, multiple ecosystems.

AI qayerda imkon beruvchi (asosiy emas)

• Pro dashboard, content recommendations, AI-powered insights.
• Push management [planned], drip campaigns uchun auto-generated copy.
• @appss search [planned], semantic search.

Bular foydali AI ilovalari. Biz «AI moat» deganda aynan buni nazarda tutmaymiz.

To'liq AI stack

Provayder	Biz foydalanadigan modellar	Kim tomonidan ishlatiladi
Anthropic Claude	Sonnet 4, Haiku, Haiku-vision	Designer Studio, Market Research AI, Social Tracker, Builder API
OpenAI	GPT-4o (fallback), gpt-image-1 (edit), Whisper	Market Research, Designer Studio, Social Tracker
recraft.ai	Vectorizer	Designer Studio
Replicate	rembg, [planned] SAM2	Designer Studio
quiver.ai	Text-to-vector	Designer Studio

Nima uchun bu moat

1. Compositional moat. Individual providers hamma uchun mavjud. Ularning barchasini bitta Telegram-creator-flow ostida boshqarishimiz esa emas. O'n ikki imkoniyatni bitta user-flow atrofida muvofiqlashtirish xarajati haftalarda emas, oylarda to'lanadi.

2. Data moat. Market Research'dagi Claude agents 2+ yil davomida to'plangan bizning 121-jadvalli production database'imizni o'qiydi. Xuddi shu modellar va tools'ga ega raqobatchida data layer yo'q. Same SQL, different answers.

3. Cost optimisation configuration emas, balki engineering sifatida. 3-tier Social Tracker, Designer Studio'dagi recursion-stop heuristics, uzoq davom etadigan agent sessions bo'ylab prompt-caching, bular flag flips emas, balki engineered details. $/operation uchun poyga oylar davomida sayqallagan jamoalar tomonidan yutib olinadi.

4. Fallback resilience. Claude → GPT-4o; recraft → Replicate. Biz bitta provider'ning uptime'iga bog'liq emasmiz.

AI-driven UX patterns biz foydalanadigan

• Candidate model. AI candidate-version'ni yaratadi; user «Apply» tugmasini bosib tasdiqlaydi. Hech qachon «AI jim turib narsalarni o'zgartiradi» emas.
• Multi-modal recursion. Agent tool-call natijasini kuzatadi va chuqurroq o'rganishni hal qiladi.
• Style descriptor passing. Bitta tool keyingisi uchun tayyor structured-string'ni qaytaradi.
• Fallback ladder. Step 1 → Step 2 → Step 3 (foydalanuvchidan so'rash). Hech qachon «AI jim turib so'ralgan narsadan farqli narsa yaratdi» emas.
• Single answer, per-surface renderer. Kanonik javob og'ir model tomonidan bir marta ishlab chiqariladi. Har bir UI surface (web, Telegram bot, kelajakdagi Discord / LINE) uni o'sha surface'ning cheklovlariga moslashtirilgan arzon ikkinchi parser-prompt (Gemini-Flash / Haiku) orqali qayta ko'rsatadi. Cross-ecosystem expansion shunday qilib arzon bo'lib qoladi, same answer, new renderer per surface.

Raqamlar va iqtibos qilinadigan da'volar

• 6 AI / ML providers integratsiya qilingan.
• Designer Studio'da 14 atomic tools, agent loop'da 10 max iterations.
• Social Tracker'dagi 3-tier pipeline = naive heavy-model-everywhere'ga nisbatan ~10× cost reduction.
• Market Research AI tomonidan so'raladigan 121 production tables + 60+ event types.

Ochiq savollar

• Oylik umumiy AI spend.
• Har bir user uchun cost (Pro / Free / pay-as-you-go).
• Fallback hit-rate (Claude → GPT-4o).
• O'rtacha prompt-caching hit-rate.

Keyingi o'qish

• Positioning, nima uchun tool emas, balki stack bo'lish bu orchestration'ni product'ga aylantiradi.
• Business model, Market Research AI qahramon use-case sifatida pricing'ni qanday shakllantiradi.
• Two-sided marketplace, nima uchun xuddi shu agent infrastructure loop'ning har ikki tomonida turli qiymatlarni ochadi.

Umumiy spend, har bir user uchun cost va cost-optimisation roadmap uchun, mark@engagelabs.org manziliga email yuboring.