從數據
到決策 From Data
to Decisions

影視與短劇行業的核心成本在於前期製作：分鏡、美術與視覺方案往往佔據數週工期與大量人力。MOPO Media 將這一環節產品化，使一部作品的視覺方案能在極短周期內成型，並在數百個鏡頭之間維持統一的美術風格與角色形象——後者正是同類工具最難跨越的門檻，也是這條技術路線的核心壁壘。目標市場涵蓋影視製作、短劇、廣告與品牌內容。In film and short-form drama, the dominant cost sits in pre-production: storyboarding, art direction and visual development routinely consume weeks of skilled labour. MOPO Media productises that stage, so a project's visual treatment takes shape in a fraction of the time while holding a consistent art style and consistent character identity across hundreds of shots — the threshold competing tools find hardest to cross, and the core defensibility of this approach. Target markets span film production, short-form drama, advertising and brand content.

達人營銷是跨境電商增長最快的渠道，也是最依賴人力的環節——尋找、篩選、建聯、議價、寄樣、追蹤成效，每一步都需要人手重複執行，規模因此受限於團隊人數。本系統以多智能體架構全天候承接整條鏈路，使單人可管理的達人規模提升一個量級。所有對外動作先進入人工審批隊列，自動化邊界由配置明確劃定——這是同類系統進入品牌客戶所必須具備的可控性。Creator marketing is the fastest-growing channel in cross-border commerce and also its most labour-bound: sourcing, screening, outreach, negotiation, sampling and performance tracking are each executed by hand, so scale is capped by headcount. This system carries the entire chain around the clock on a multi-agent architecture, raising the creator volume one person can manage by an order of magnitude. Every outbound action enters a human approval queue first, with the boundary of automation set explicitly in configuration — the control brand clients require before adopting a system like this.

人口老化是香港與全球共同面對的結構性議題，而長者照護中最難自動化的一環，是在「看似無事」時確認真的無事。一撳平安以一次日常確認為起點：長者每日撳一次「我平安」，未確認時系統按聯絡人次序逐級升級，直到有人應承跟進為止。這類系統的技術重心與生成式應用恰好相反——它要求的不是模型的創造力，而是狀態機的絕對可預期性：升級路徑由確定性規則驅動，並以自動化回歸測試逐條鎖定，確保系統不會在看起來正常的狀態下靜默失效。判斷一項業務應由模型承擔、還是應由確定性工程承擔，是我們在諮詢工作中反覆面對的核心問題；這個產品是我們對後者的完整實踐。Population ageing is a structural challenge for Hong Kong and for the world, and the hardest thing to automate in elder care is confirming that nothing has happened when nothing appears to have happened. StillOK begins from a single daily confirmation: the senior taps “I’m OK” once a day, and if that confirmation does not arrive the system escalates through the contact list until someone acknowledges. The engineering centre of gravity here is the inverse of a generative application — what it demands is not a model’s creativity but a state machine’s absolute predictability: escalation paths run on deterministic rules, pinned by automated regression tests, so the system cannot fail silently while appearing to work. Judging which parts of a business belong to a model and which belong to deterministic engineering is a question we work through constantly in advisory engagements; this product is our fullest expression of the latter.

全球中文學習市場持續擴張，但絕大多數產品止於認讀，真正涵蓋「寫」的極少。原因在技術：選擇題只需比對答案，而書寫要求系統即時判斷每一筆的起止、方向與次序，並在孩子落筆的當下給出可理解的反饋——這是一個識別精度與推理延遲同時受約束的模型問題。PandaPal 把讀、聽、寫做進同一條學習路徑，以手寫識別與語音合成支撐逐筆反饋，並支持簡繁雙軌，面向兩岸三地與海外華人家庭。兒童教育對錯誤的容忍度極低：模型判錯一筆，孩子學到的就是錯的寫法。如何在兒童設備可接受的算力下，把識別精度推到足以用於教學的水平，是這條產品線的核心工程課題，也是同類產品長期停留在認讀階段的原因。The global market for Chinese language learning keeps expanding, yet almost every product stops at recognition and very few genuinely cover writing. The reason is technical: multiple choice only has to compare an answer, whereas writing requires the system to judge the start, direction and order of every stroke as the child forms it, and to return feedback the child can act on — a model problem constrained simultaneously by recognition accuracy and inference latency. PandaPal brings reading, listening and writing into one learning path, using handwriting recognition and speech synthesis to drive per-stroke feedback, with switchable simplified and traditional characters for Greater China and overseas Chinese-speaking families. Children’s education tolerates error poorly: if the model misreads a stroke, the child learns the wrong form. Pushing recognition accuracy to a teachable standard within the compute budget of a child’s device is the central engineering problem of this line — and the reason comparable products have stayed at recognition.

國際象棋是全球滲透率最高的兒童素質教育品類之一，但入門長期依賴人工授課。棋星小課堂把完整規則體系拆解為五十六個循序漸進的關卡，並配備一位由大型語言模型驅動的「AI 軍師」。難點不在接上模型，而在讓它不說錯棋——本產品把「算」與「講」拆開：棋局事實一律由自研引擎給出（Negamax 與 Alpha-Beta 剪枝、靜態交換評估、迭代加深，並通過完整 FIDE 規則與 perft 驗證），模型只負責把引擎的結論翻譯成初學者聽得懂的語言，不得自行推算。這一套「引擎接地」的分工，是把大型語言模型放進棋類任務而不產生幻覺的關鍵。介面涵蓋繁體中文（粵語）、簡體中文、英文與西班牙文，現已在 Google Play 上架，iOS 版本即將推出。Chess is among the most widely adopted enrichment categories for children worldwide, yet entry has long depended on human instruction. Chess Stars decomposes the full rule set into fifty-six progressive levels alongside an LLM-driven tutor. The difficulty is not connecting a model but stopping it from suggesting illegal or imaginary moves, so the product separates calculation from explanation: every board fact comes from an in-house engine (Negamax with alpha-beta pruning, static exchange evaluation and iterative deepening, verified against the full FIDE rule set and perft), while the model only renders the engine’s conclusions into language a beginner understands, and is barred from reasoning about positions itself. That grounding is what makes a language model usable in a chess task at all. The interface covers Traditional Chinese (Cantonese), Simplified Chinese, English and Spanish. It is available on Google Play, with the iOS edition to follow.