Skip to content
··閱讀時間3分鐘

STRAŦUM:我75日內Solo Build嘅9-Agent Marketing Application(其中病咗10日)

我75日內build咗一個9-agent marketing platform,佢從每段對話中學習——同一個agent講你嘅business,九個全部都會變得更smart。

記得9月嗰篇文我casual咁提過一邊瞓Sunday午覺一邊speed-run一個10-agent marketing platform?四個禮拜入面,3個agents working,target 10月做alpha?

同埋10月嗰個debugging噩夢我reveal咗個名STRAŦUM同提到有8個(9個入面)agents build好?

好喇,依家11月喇。係時候actually launch呢件嘢。

狀態:

- ✅ Platform名:STRAŦUM(Intelligence Over Execution)

- ✅ 9個(9個入面)AI agents build好同integrated

- ✅ SMEs同agencies嘅Multi-tenant architecture

- ✅ 完整brand guidelines同design system

- ✅ Private-alpha testing phase:而家

Marketing Execution冇Strategy只係貴嘅Noise

大部分platforms promise更快execution。但冇方向嘅速度只係burn budget更快。

STRAŦUM唔同:11個strategic frameworks。9個AI agents。隨住每段對話增長嘅intelligence。

以下係佢actually嘅意思:

11個Strategic Frameworks應用到你嘅Business:

SWOT Analysis、Porter's Five Forces、Blue Ocean Strategy、BCG Matrix、VRIO、McKinsey 7S、OKRs、Three Horizons、ICE Prioritization、Business Model Canvas、Jobs to Be Done。

Progressive Learning - 同一個Agent講,九個全部都知:

唔似static tools,STRAŦUM嘅AI agents從每段對話中持續學習你嘅business。你interact得越多,insights越smart越targeted。

點運作:

第1日 - 自動Insight Capture:

Share你嘅business context一次。Agents提供strategic frameworks同開始自動capture關於你嘅market、competitors同customers嘅key insights。唔使撳「Save」——intelligence自然累積。

第1個禮拜 - Cross-Agent Intelligence Sharing:

每段對話都加入你嘅Learning History。同Strategy Agent傾European expansion?Content Agent下個禮拜已經知。Agents reference過去嘅insights提供越來越personalized嘅recommendations。唔使再解釋context。

持續 - Predictive Intelligence:

有咗rich knowledge base,agents anticipate你嘅需要同pre-fill contexts。高信心insights(≥90%)自動approved。低信心learnings等你review。你保持控制——view你嘅完整Learning History同delete你唔鍾意嘅任何insight。

5分鐘Quick Wins:

第一次session就有actionable insights,唔係第五個禮拜。

Evolution:由DIALØGUE到STRAŦUM

Build DIALØGUE教識我點ship AI products。Build STRAŦUM教識我點build platforms。

DIALØGUE(8個月,Jan-Aug 2025):

- 一種user type:Individual podcast creators

- 一個workflow:Research → Script → Audio generation

- 14 microservices:Lambda → Cloud Run migration

- Simple auth:一種user type嘅JWT

- 一個revenue stream:Credit packs($4.99-$19.99)

- Business model:B2C、single-tenant

STRAŦUM(75日,Aug-Nov 2025):

- 兩種user types:SMEs + Agencies(各manage 5-15個clients)

- 9個specialized agents:每個有multiple tools,share intelligence

- 45+ database tables:完整multi-tenant data isolation

- Complex auth:Organization → Client → Campaign hierarchy

- Business model:B2B + B2C、multi-tenant SaaS

點解Multi-Tenant Architecture難10倍

DIALØGUE一次為一個user generate podcasts。STRAŦUM manage marketing intelligence for agencies juggling multiple competing clients。

挑戰:一間manage多個clients嘅agency需要:

- 完整data separation(Client 1睇唔到Client 2 data)

- Hierarchical context(organization → client → campaign)

- Cross-agent intelligence sharing(只喺campaign boundaries內)

- Schema routingpublic schema for SME、agency schema for agencies)

- Row Level Security on 45+ database tables

例子:當strategist為Client 1用Business Strategy Agent,嗰個analysis係isolated嘅。Switch到Persona Agent?佢pull Client 1嘅strategy——但_睇唔到_Client 2。Brand guidelines cascade down到Content Agent,確保每篇content用Client 1嘅voice,唔係Client 2嘅。

呢個需要database functions、materialized views、trigger-based real-time updates,同老實講比我想寫嘅更多SQL。

200+ commits over 72日淨係for multi-tenant architecture。嗰個唔係feature——嗰個係architectural philosophy,touch咗stack嘅每一層。

數字(因為我忍唔住)

75日。8月20日到11月3日。以下係需要嘅:

Development Velocity:

- Git commits:1,000+(actual count: 1,075)——嗰個係平均每日14.5個commits

- Lines of code:~200,000(Python: 62k、TypeScript: 98k、SQL: 41k)

- Database migrations:214個sequential migrations

- Agents built:9個(9個入面)(所有core agents shipped)

Technical Complexity:

- Database tables:45+ tables有完整RLS policies

- RLS policies:83個policies across 26個tables for multi-tenant security

- Foreign key indexes:98個indexes added(Postgres唔會auto-create佢哋!)

- Color token migration:700+個instances across 200+ files一日內

- Major architectural pivots:3個(ADK→Direct API、Nuclear Migration、Database-First)

Performance Improvements:

- Latency reduction:72% faster AI responses(hybrid function calling)

- RLS optimization:10-100x query speedup有policy caching

- Bundle size:92% reduction through code splitting

Reality Check:

- Multi-tenancy嘅navigation bugs:23個(2日fix完)

- 病假日數:10日(仲係準時ship)

- 放假日數:8日(beach ≠ debugging)

- 飲咗幾多咖啡:仲係唔好問

- 差啲放棄嘅次數:0 :P

- Claude Code / Gemini 2.5 Pro救咗我幾多次:老實講數唔晒

我(又)學到咩

1. Multi-Tenancy好難

Data isolation唔淨止係加org_id到每個table。要think through:

- 呢個data住喺邊個schema?(public for SME、agency for agencies)

- 當你delete一個campaign會發生咩?(Soft delete用archived_at,唔係hard delete)

- Permissions點cascade?(Organization admin vs client manager vs campaign contributor)

例子:11月1日,我一日fix咗23個navigation bugs。問題?Agency users navigate between clients break URL context。SME routes好似/persona/session/123,但agency routes需要/clients/[client-slug]/agents/persona/session/123。每個agent page都需要refactoring嚟preserve client context across navigation。

嗰個係200+ commits over 72日淨係for multi-tenant architecture。唔係feature——係architectural philosophy,touch咗stack嘅每一層。

2. 10日病假Break Momentum

我提到嘅10月launch?係,我病咗。睇唔到screen。寫唔到code。只能...等。

Solo development代表冇team pick up slack。但佢亦代表冇壓力逼你喺ready之前ship。我choose咗get it right over getting it fast。

3. AI-Assisted Development係Real嘅(但唔係Magic)

9月嗰個speed-run唔係誇張。Claude Code同Gemini CLI令我可以幾個鐘內ship architectural refactors,以前要幾日。

9月14日嘅例子:

```

08:04 AM - Migrated frontend to standardized API client

11:34 AM - Centralized route configuration (no hardcoded URLs)

1:00 PM - Standardized all 10 agent pages

4:38 PM - All agents integrated with context system

5:03 PM - Testing & Polish (92% bundle size reduction)

```

六個major features。一個Sunday。期間去教堂、買餸、食lunch、瞓午覺、用iPad打game。

但係:AI冇寫architecture。佢冇決定multi-tenancy patterns。佢冇debug嗰個HTTP/HTTPS噩夢。佢_amplified_我嘅decisions。Thought partner,唔係ghostwriter。呢個lesson喺我開始build一個native iOS app但唔識Swift嘅時候變得更sharp——Claude Code一個晚上scaffold咗7,568行,但product taste decisions仍然係我嘅。

Shape咗STRAŦUM嘅3個Pivots

Build呢個platform需要三個major architectural decisions,fundamentally改變咗trajectory:

Pivot 1:Direct Gemini API(第2日 - 8月21日)

第2日,我abandon咗Google嘅ADK(佢哋嘅Gemini SDK wrapper)轉用direct API access。ADK有session management limitations同multi-tenant architecture conflict。Migration用咗24個鐘。Early pivots便宜。Late pivots貴。

Pivot 2:由第1日起Multi-Tenant

我可以淨係build for SMEs。相反,我choose support agencies managing multiple clients。呢個decision加咗3個月嘅complexity:schema routing、data isolation、client context propagation。但佢亦開啟咗enterprise sales potential——agencies managing 5-15個clients proportionally pay更多。每個customer 10倍revenue justify 3倍development time。

Pivot 3:Nuclear Migration(10月11-22日)

到10月,我有9個separate intelligence tables(每個agent type一個)。每個新agent需要新migrations、新API endpoints、新frontend queries。我consolidate咗全部9個入一個unified table有flexible schema-less content。Migration用咗11日。而家加新agents用幾個鐘,唔係幾日。

呢啲唔係technical failures——佢哋係strategic decisions。AI幫我execute得更快,但architecture decisions係我嘅。

咩Work緊

- 9個AI marketing agents應用11個strategic frameworks

- Multi-tenant data isolation——agencies可以safely manage multiple clients

- Progressive learning system——cross-agent intelligence sharing within campaigns

- Real-time SSE streaming for所有agent conversations

- Interactive persona interviews capture nuanced customer insights

- Marketing strategy bridging business strategy到tactical execution

呢個係working platform,唔係vaporware。Private alpha代表real users已經喺test緊。

Private Alpha:申請Early Access

STRAŦUM live喇,透過invitation-only access接受early testers。我搵緊:

- Small businesses或startups(1-10人)需要strategic marketing intelligence

- Marketing agencies managing multiple clients想要efficient strategy tools

- Early adopters想透過feedback shape product

呢度申請access。我personally review每個request,24-48個鐘內grant access。

你會得到:

- 9個AI marketing agents有11個strategic frameworks

- Multi-campaign management(agencies:manage multiple clients)

- Progressive learning system隨住每段對話變得更smart

- 直接contact我做feedback同feature requests

預期:

- Private alpha = actively evolving based on user feedback

- 我對bugs同feature requests好responsive

- Solo founder = authentic、hands-on support

更大嘅Vision

做咗20年advertising之後,我見過同一個pattern:好嘅marketing strategy好貴同難access。Agencies收五位數每個月。好嘅strategists年薪六位數。Solo founders同small teams被遺漏。

但如果strategic marketing intelligence可以被AI augment_呢?唔係replace——augment。AI handle frameworks、research、structured thinking,而人類帶嚟creativity、intuition、嗰種令marketing actually work嘅_je ne sais quoi

嗰個就係STRAŦUM。Intelligence over execution。Strategy over tactics。Thought partner over ghostwriter。

佢完美嗎?Hell no。佢有用嗎?我genuinely覺得係。

最後諗法(或者:點解我繼續Build in Public)

Build STRAŦUM比DIALØGUE更難。更複雜。更貴。更多夜晚question緊有冇人actually想要呢個。

但以下係令我繼續嘅嘢:同我build DIALØGUE嘅原因一樣。因為我想要佢。

同因為document呢個journey——victories、debugging噩夢、23個navigation bugs、3個architectural pivots、214個database migrations——幫其他solo builders睇到咩係possible嘅。

75日前,一個人build一個9-agent marketing platform有multi-tenant architecture好似impossible。今日,我invite緊人test佢。

咩變咗?更好嘅AI tools——尤其係Claude Code。

你有冇solo build緊ambitious嘅嘢?或者考慮緊?我好想聽你做緊咩同咩hold you back。由我嘅經驗嚟講,最難嘅部分唔係code——係decide to start。

祝好,

Chandler

想試STRAŦUM?申請invitation。

繼續閱讀

我嘅旅程
聯繫
語言
偏好設定