第4個禮拜：Build一個10-Agent Marketing Platform同時仲瞓午覺
我喺4個禮拜內build咗3個AI marketing agents加multi-tenant architecture——同我之前花咗7個月先launch嘅product第一個月嘅進度一樣多。
更新（2025年11月）： STRAŦUM依家live喇！由一個「4-week speed run」開始演變成一個完整嘅9-agent marketing intelligence application。睇完整launch故事：STRAŦUM：我75日內Solo Build嘅9-Agent Marketing Application
**即刻試：**去stratum.chandlernguyen.com（Private Alpha - 只限邀請）
六日。呢個係我launch咗DIALØGUE之後等咗幾耐先開始我嘅下一個project。
你可能會諗「你唔應該，我唔知，做marketing你嘅podcast generator？搵用戶？修bugs？」而你_完全啱_。但係咁嘅——我想睇吓我可以用新AI工具push自己行得幾遠幾快。叫佢speed run也好、過度ambitious也好、叫佢當Claude Code同Gemini 2.5 Pro令build變返_好玩_嘅時候會發生咩也好。
唔應該可能嘅Timeline
等我put this in perspective：
- DIALØGUE：由第一行code到launch用咗6-7個月
- Marketing Suite：4個禮拜入面，已經有3個working AI agents互相傾偈
睇git history，第一日結束時，我有：
- Multi-tenant architecture，有organization → client → campaign hierarchy
- Initial React frontend配authentication
- 一個working Business Strategy Agent用11個唔同frameworks，涵蓋唔同regions。呢啲frameworks係top global/regional consultancy firms用嘅。例如：
- Supabase integration配Row Level Security
呢個比我build DIALØGUE頭一個_月_做到嘅仲多。
令我繼續嘅數字
過去4個禮拜嘅quick metrics：
- Features shipped：3個完整agents + multi-tenant architecture
- Lines of code：~117,000（Python: 39k、TypeScript: 58k、SQL: 20k、Javascript: 其餘）
- Database tables：39（每個都serve特定用途——稍後詳述）
- Git commits：232（大約26日）
- 飲咗幾多咖啡：唔好問
點解呢個Project實際上難10倍
DIALØGUE複雜，冇錯。AWS Lambda functions、Step Functions，最終migrate所有嘢到Google Cloud Run。但fundamentally，佢係一個single-purpose工具：generate podcasts。一種user type。一個workflow。一個happy path。
呢個Marketing Suite？等我show你我嘅意思：
一個真實場景： 一間agency manage Nike同Adidas（hypothetically）。Agency嘅strategist用Business Strategy Agent分析Nike嘅position。嗰個analysis自動save到strategy_outputs table。當佢哋switch到Persona Agent，佢pull Nike嘅strategy嚟inform persona development——但佢_睇唔到_Adidas嘅data。同時，Nike嘅brand guidelines喺brand_guidelines table入面cascade down到Content Agent，確保每篇content用Nike嘅voice，唔係Adidas嘅。
呢個需要：
- organizations table有SME vs AGENCY types
- clients table（只限agencies）
- campaigns有foreign keys到org同client
- strategy_outputs有campaign-level isolation
- brand_guidelines有hierarchical inheritance
- 每個table都有Row Level Security policies
呢個只係_一個_workflow。而家乘以10個agents，每個都有佢哋自己嘅data requirements。
咩Actually Work緊（同咩仲係Chaos）
Work緊嘅（3個Live Agents + Business Intelligence）
Business Strategy Agent：呢個唔係你嘅typical SWOT generator。佢apply 11個comprehensive frameworks：
- SWOT Analysis
- Porter's Five Forces（Competitive rivalry、Supplier power、Buyer power、Threat of substitution、Threat of new entry）
- Business Model Canvas（9 building blocks of business design）
- ICE Prioritization（Impact、Confidence、Ease scoring）
- BCG Growth-Share Matrix（Stars、Cash Cows、Question Marks、Dogs）
- VRIO Framework（Value、Rarity、Imitability、Organization）
- Three Horizons Model（Current core、Emerging opportunities、Future bets）
- Blue Ocean Strategy（Eliminate、Reduce、Raise、Create grid）
- McKinsey 7S Framework（Strategy、Structure、Systems、Shared Values、Skills、Style、Staff）
- OKRs Framework（Objectives and Key Results）
- Jobs to Be Done（Customer jobs、Pains、Gains）
每個framework generate嘅structured data feed入其他agents。當你run一個SWOT analysis，「Opportunities」自動inform Marketing Strategy Agent嘅growth tactics。
Persona Agent：Generate detailed customer personas有15+個attributes而且——呢個係wild嘅部分——你可以interview佢哋。Actual conversation：
You: "What frustrates you most about current project management tools?"
Persona (Tech Startup Founder): "The constant context switching. I need to check Slack, then Asana, then our analytics dashboard. By the time I figure out what needs attention, I've lost 30 minutes."
每個interview response會越嚟越smart因為佢pull from persona_interactions history。用戶花20-30分鐘做呢啲interviews。一個用戶話「It's like focus groups but instant and actually useful.」
Marketing Strategy Agent：呢個係bridge。佢攞你嘅business strategy、understand你嘅personas、然後create actionable go-to-market plans。佢唔只係話「use social media」——佢map specific personas到specific channels配specific messaging。Bootstrapped startups嘅zero-budget tactics。SMEs嘅70-20-10 budget allocation。Enterprises嘅multi-channel orchestration。
Business Intelligence（Hidden Hero）：同任何agent嘅每段對話都會自動extract insights同build organizational knowledge。同Strategy Agent傾進入European market？嗰個insight save到ai_insights table。下個禮拜你用Content Agent嘅時候，佢已經知你嘅European expansion plans。唔使再解釋。冇context loss。
仲喺Build緊嘅（下一批7個Agents）
Campaign Execution Agent：呢度啲嘢開始scary。你點樣safely store Google Ads、Meta、LinkedIn、TikTok嘅API keys？Current plan涉及encrypted storage喺platform_credentials table入面配audit logs for every API call。但permission model令我夜晚瞓唔著。
Analytics Agent：先integrate邊個APIs？Google Analytics 4？Meta ads？LinkedIn Analytics？挑戰唔淨止係pull data——而係normalize across platforms等你可以actually比較蘋果同蘋果。
ROI & Budget Agent：跨多個平台嘅real-time tracking配唔同attribution models。你點reconcile Google嘅data-driven attribution同Meta嘅model？仲喺度figure out。
其他4個Agents：Quick Wins（搵即時機會）、Competitive Intelligence（ethical competitor analysis）、Client Success（retention strategies）、同Content Agent（actually已經work但需要polish）。
令我夜晚瞓唔著嘅難題
- Brand name：我應該點樣叫呢個application？:P 如果你有好suggestions，話俾我知！
- API Security：你點handle multiple ad platforms嘅API keys而唔變成security nightmare？
- Cohesive Experience：你點make 10個specialized agents feel like一個unified platform而唔係10個唔同tools用膠紙黏埋？
- 同好多好多 :)
Architecture Evolution：由Pain中學習
用DIALØGUE，我hard way學到race conditions。記得嗰個3分鐘signup bug嗎？新用戶要等forever因為auth trigger同Edge Function racing嚟create同一個user record。
呢次，我由第一日開始就build啱。以下係幾日前嘅actual example：
問題： Marketing Strategy Agent需要知道存在咩personas、create咗咩strategies、同follow咩brand guidelines——全部同時maintain complete data isolation。
Solution： 唔係每個agent query多個tables（慢、複雜、error-prone），我build咗一個Enterprise Context Service做single source of truth：
// Before: Each agent making multiple queries
const personas = await supabase.from('personas').select()
const strategies = await supabase.from('strategies').select()
const guidelines = await supabase.from('brand_guidelines').select()
// After: One intelligent service
const context = await getEnterpriseContext(campaignId)
// Returns filtered, cached, properly-scoped data in 45ms
結果？Content Agent可以即刻access brand voice guidelines、睇到邊個personas需要為佢哋寫、同understand marketing strategy——全部喺一個respect data boundaries嘅call入面。
速度差異係Real嘅
9月14日係insane嘅。睇actual git log配timestamps：
08:04 AM - Migrated frontend to standardized API client
08:19 AM - Completed API standardization (100% coverage)
11:34 AM - Phase 1: Centralized route configuration
(no more hardcoded URLs anywhere)
1:00 PM - Phase 2: Standardized all 10 agent pages
(consistent URL patterns like /agents/strategy/:campaignId)
1:10 PM - Phase 3.1: Core context management
(workspace → client → campaign hierarchy in URLs)
4:31 PM - Phase 3.2: Navigation components became context-aware
(breadcrumbs show "Nike › Summer Campaign › Strategy Agent")
4:38 PM - Phase 3.3: All agents integrated with context system
(switching campaigns preserves your place)
5:03 PM - Phase 3.4: Testing & Polish
(92% bundle size reduction through code splitting)
六個major architectural improvements。淨係URL system就touch咗40+個files across codebase。冇AI assistance嘅話，呢個會係一個禮拜嘅refactor，中間可能break幾個features。
而且我其實大部分時間唔係喺電腦前面 :D
嗰日係Sunday所以我哋去咗教堂、超市、食咗一頓有海鮮嘅full lunch、瞓咗午覺、去咗Costco同用iPad打咗啲game。呢個係AI assisted coding令到可能嘅。多謝Claude Code同Gemini CLI！
以下係URL standardization actually對用戶嘅意義：
// Before: Lost context when switching between agents
"/strategy" // Which campaign? Which client? Who knows?
// After: Context preserved in URL
"/workspace/nike/campaign/summer-2025/agents/strategy"
// Bookmark it, share it, refresh it - context stays intact
點解DIALØGUE啱Launch就Build呢個？
因為我可以。因為啲工具已經_好到咁_。因為做咗20+年advertising之後，我終於有skills build到我一直想用嘅marketing platform。
但主要？因為我想document當你combine domain expertise同modern AI tools嘅時候咩係possible嘅。呢個唔係關於replace developers——而係amplify佢哋。四個禮拜前，build咁複雜嘅嘢需要一個team。而家佢需要determination、好嘅AI assistants、同dangerous amount嘅caffeine。（雖然我之後喺唔識Swift嘅情況下build一個native iOS app時發現，AI俾你大概60%——最後40%嘅taste同polish仍然完全係人類嘅。）
下一步？
我target 10月/11月做alpha launch。以下係要做嘅：
- Finish其餘7個agents
- Build campaign planning interface
- Figure out ad platform integration security
- 確保agency UI/flow sensible
- 用actual users test所有嘢（terrifying）
淨係Analytics Agent就需要aggregate Google Analytics、ad platforms同internal metrics嘅data入unified dashboards。ROI Agent需要near real-time budget tracking across multiple platforms。
每一個本身就係一個project。但以呢個pace？佢actually感覺係possible嘅???
想follow along？隨住我build我會share更多updates。
Build in public代表share chaos同victories一齊。而家大部分係chaos。但佢係_productive_ chaos，呢個先係count嘅。
一個月後再返嚟。如果一切順利，你可以test到一個marketing platform，有AI agents做到成個marketing department嘅工作。如果唔順利...至少啲blog posts會好entertaining。:P
