Case · AI orchestrationin production

Not a team and a quarter, but a pipeline and a week: a turnkey digital ecosystem

For an expert partner I designed and built an app, a book, a Telegram bot, three landing pages and a publishing engine — six connected products in one funnel, through AI orchestration, in 10 days.

6
products in the funnel
10
calendar days
145
commits
200+
AI agents

доказаноevery number checkable in code and gitкак проверяем →

How it all began

A warm audience with no product — into a funnel of six

The expert partner had years of hard-won expertise in women’s fitness and a warm audience that trusts her. There was nothing that turns that trust into a product: no engineering team, no app, no funnel from free content to payment. The task — turn it into an honestly selling ecosystem.

Going in

An 18-sheet program in Excel

The expertise and the warm audience are there. An engineering team, a sellable product and a funnel that turns reach into revenue — are not.

Coming out · 10 days

App · book · bot · 3 landing pages · publishing engine · funnel

Six subsystems glued into one funnel. Designed, built, verified, shipped to production.

Three constraints shaped the whole architecture: honesty instead of the market’s dark patterns, safety on a sensitive topic (a route to a doctor, not a protocol) and data-protection law built into the design, not bolted on before launch.

What was built

Six products, one funnel

Six subsystems — each a product in its own right, but all glued into one shared customer path. Next to each — its role in the funnel and a scale you can check in the code.

Three landing pages — reach and conversion

Three product entry points in one repository: 8 routes, all statically prerendered; interactivity is targeted client islands, not full hydration of the page. Transparent prices, 0 fake discounts.

olgakashina.ru ↗

The book “Biology of a Winner” — the trial product

The trial product (₽880): 43 chapters / 261 A5 pages, PDF + reflowable EPUB 3 (epubcheck — 0 messages), 3 editions from a single source, 59 named cross-references.

book.olgakashina.ru ↗

The PWA “Dream Glutes” — the core

The program product (₽2490): an 8-week strength program, nutrition, a progress section, offline sync. 42,510 lines of TS/TSX, 111 components, 21 REST routes, 25 tables, 66 tests.

app.olgakashina.ru ↗olgakashina.ru/app ↗

The Telegram bot — top of the funnel

A free 7-day drip series, 16 guaranteed touches, consent before the message. 31 files, 1543 lines, a 7-state state machine. It hands payment off to the app’s checkout — and owns the rhythm itself.

t.me/bootyfit_bot ↗

The content machine — strategy and reach

Competitive intelligence: 9 competitors, 785 links, 36 landing pages, 42 verified prices. A unified brand strategy and an ethical frame: for every dark pattern, an honest replacement was found.

t.me/kashina_fit ↗vk.com/kashina_fit ↗

Design system and infrastructure

An agent-callable design system (tokens with no hardcoded hex colours), a video-production tool, a meta-repository with CI and document memory: 1 head + 6 nested modules, a single “ecosystem version” commit.

End-to-end funnel

Five stages — from reach to retention

The six products aren’t a set, they’re a route. A subscriber enters at the top on free content and descends to the core program. The principle running through every stage — honesty instead of dark patterns: transparent prices, hooks with no email harvest, a route to a doctor instead of fearmongering.

1

Reach

no email

A link in the profile plus a hook and a sample chapter. A model of attraction: reciprocity with no strings.

2

Free 7-day drip series

free

Telegram bot: 16 touches (morning / evening / check-in), consent before the start, hand-off to paid on day 7.

3

Trial product — the book

₽880

“Biology of a Winner” in PDF + EPUB. A landing page with a preview, payment, and an access email.

4

Core — the program

₽2490

A PWA app: sign-in (PAR-Q + consent) → workouts / nutrition / progress.

5

Retention

Lifecycle notifications in Telegram via an outbound queue, reminders, a small reward for the path completed.

The prices ₽880 and ₽2490 are the expert partner’s product prices in the funnel, not temadev’s service fees.

Engineering maturity

Behind the speed — senior-engineer decisions

Speed is suspicious — so it matters more to show what’s behind it. This is no trick and no “generate and forget”: under the hood are senior-engineer decisions, each checkable in the code. Here are eight that tell engineering from a prototype.

Offline at the core

localStorage is the source of truth; conflicts resolve by “last write wins”, the server is secondary. A deliberate choice for reliability on mobile with no network.

Payments independent of the provider

A registry maps a provider ID → adapter; providers switch by an environment flag with no single code change. Idempotency via order-id and a re-check request.

Data-protection law: two-way erasure

app↔bot through a reliable outbound queue with reconciliation. Erasure is atomic in one transaction — no break if it fails between “deleted” and “queued”.

Least-privilege tokens

Each channel gets its own secret; a missing secret = the feature is honestly switched off (503), not a crashed process. A safe, staged launch.

Adversarial checking at scale

Each block passes N independent veto lenses plus a 4th layer — a fresh agent with a mandate to “break this”. Not a sequential review, but real adversarial pressure.

One source → many editions

One stream of book tokens → PDF + EPUB + a short summary; one product config → landing pages and offers. Edit in one place, deterministic rebuild — the author touches only .md.

Static at the core, targeted interactivity

Every route is statically prerendered; client JS only where interactivity is needed: targeted islands, not full hydration of each page. “‘use client’ when needed”, not by default.

Health safety as determinism

Sensitive input is routed by fixed regex patterns, not free LLM generation. On risky topics — a route to a doctor, not a prescribed protocol.

Method · AI orchestration

Not “I use ChatGPT”, but an adversarial pipeline

How do you fit so much into 10 days? The lever isn’t typing speed but a designed orchestration pipeline: agents research, write and assemble, while I hold the truth and the decisions. The key — checking that isn’t sequential but adversarial: every block is attacked, not just “looked at”.

  1. step 1
    Draft

    an agent writes a block

  2. 3 veto lenses ∥
    Attack, not review

    facts · safety · voice — in parallel

  3. decision
    To production or back

    a binary quality gate

  4. 4th layer
    “Break this”

    a fresh agent hits the assembled block

One adversarial run — the book audit: 62 agents → 77 findings → 53 confirmed, 1 false. 200+ agents ran through the content cycles, each with a real wf_* id.

доказаноthe audit and the agent count — by reports with wf_* idscheck on /proof →

Honest status

Live. I claim only what’s provable.

The system is shipped to production, YooKassa payments work — the products can be opened right now. I don’t speak of revenue or customers as fact: the first cohort has only just started. I claim only what’s provable — I’ll publish the numbers once I can confirm them.

  • in production

    YooKassa payments work. The provider switches by a flag: lava.top ↔ YooKassa — with no code rewrite.

  • in production

    Data protection by law: consent on collection and full erasure on request — in both the app and the bot — built into the code, not bolted on before launch.

  • in production

    42,510 lines of TS/TSX, 66 tests, CI gates on every module, EPUB epubcheck — 0 messages.

  • awaiting the cohort

    I don’t pass revenue or customers off as fact — the first cohort has only just started. I claim only what’s provable; I’ll publish the numbers once I can confirm them.

  • awaiting the cohort

    Reviews and before/after — I publish only with consent and only once I can confirm them.

What this proves

A designed pipeline — the work of a whole team

Six connected products in one funnel in 10 days and 145 commits — by orchestrating 200+ agents. Not “a person plus ChatGPT”, but a designed pipeline with adversarial checking, where every number is checkable in code — with senior-engineer discipline: CI gates, dated audits, schemas-as-contracts, debt tracked explicitly.

Next — three doors.

FAQ

Common questions

Let’s work out how this applies to your product

The same method ports to another niche. Write to me — we’ll discuss what here gives you the lever.