AI Guides
A developer's honest take on how AI actually works.
2026-05-01 · Service
Building Production AI Systems — LLM Applications, RAG Pipelines & Document Intelligence
Independent AI engineer based in Hong Kong. Production LLM applications, RAG pipelines, document intelligence. Project-based engagements.
2026-06-21
Long-Term AI Memory: The Architecture and Engineering Behind HKSoka
Seed, learned, and critical memory layers, proposition-level chunking, bilingual embeddings, hybrid retrieval — a full engineering breakdown of HKSoka's memory architecture.
2026-08-19
AI Companion Hong Kong: Three Conversation Modes That Shape How AI Responds
Three AI conversation modes — information, decision, companionship — and how to switch between them. Includes HKSoka memory system overview and Hong Kong crisis resources.
2026-08-19
Beyond AI Memory: Cognitive Models and Portable Identity
Behavioral data records what was done; cognitive memory captures why decisions were made. From event logs to decision maps — portable cognitive models as the core asset in multi-model AI work architectures.
2026-08-19
Hong Kong AI Tool Selection 2026: Platform Access, Memory Architecture, and Practical Choices
Claude access restrictions in HK, Goldman Sachs and JPMorgan bans, Perplexity-HKT partnership — 2026 platform access status, memory architecture comparison, and four selection dimensions.
2026-08-19
Autonomous AI Agents in 2026: Quantified Industry Data
Only 2% of enterprises have fully deployed AI agents; 73% of tool calls retain human presence; software engineering accounts for ~50% of agent use. Anthropic 2026 research plus OpenAI, Google, Microsoft, xAI deployment status.
2026-08-19
AI Document Processing for Hong Kong Property Management: Engineering Reality
Leases, inspection reports, maintenance records — bilingual formats, scanned PDFs, WhatsApp data flows. Which document workflows suit automation, which need human judgment, and three questions before procurement.
2026-08-19
LLM Cost Management: Real Cost Structures in Production
LLM production costs are driven by architectural decisions. Prompt caching, context window management, and multi-model strategy can reduce monthly spend to 25–30% of unoptimized levels.
2026-08-19
Fixed-Scope AI Contracts: A Buyer's Clause-by-Clause Guide
Buyer's guide to eight clauses in fixed-scope AI contracts: scope definition, deliverables, milestone payments, source code handover, API cost allocation, change request mechanisms, and warranty terms.
2026-09-25
AI-Assisted Traditional Chinese Writing: Prompts
How AI handles three Chinese writing systems, common errors such as Traditional-Simplified drift and mixed Hong Kong-Taiwan vocabulary, and four effective prompting strategies.
2026-09-25
HK Adults Learning English with AI: An EF EPI Assessment
EF EPI 2025 ranks Hong Kong 538 globally 39th, speaking is the weakest skill, AI value in grammar checking and speaking practice, pronunciation feedback limits, and where AI fits versus traditional courses.
2026-09-25
AI Contract Translation: HK Legal Limits
Hong Kong bilingual contract framework, HKCFA judgment translation research, known failure modes, 17–33% hallucination rate in legal AI, certified translation requirements and liability.
2026-09-25
AI Use Boundaries for DSE, IELTS and CFA Prep
No standalone DSE AI policy, English SBA guidance, the IELTS 2026 template crackdown, confirmed CFA curriculum stability, and effective versus risky AI uses for each exam.
2026-09-25
AI for Immigration Research: Limits and Risks
Status of BN(O), Canada, Australia and Taiwan pathways in 2025–2026, three genuinely useful uses of AI, documented failure modes and one incident, and licensing rules by destination.
2026-09-25
DeepSeek and Qwen in Hong Kong: Cost, Sovereignty
DeepSeek V4 Pro and Flash pricing, the August 2026 price rise trend, Qwen 3.6 self-hosting under Apache 2.0, data stored on Chinese servers, and local deployment.
2026-09-25
AI Search 2026: Perplexity, Google AI, SearchGPT
How Perplexity, Google AI Overviews and SearchGPT differ on citation error rate, hallucination rate, market share and Hong Kong availability, with a selection framework.
2026-09-25
Cursor vs Claude Code: Architecture and Use Cases
Cursor as IDE-driven and Claude Code as agent-driven, token and completion-time comparison, subscription pricing, known limits, and selection for daily coding versus architectural tasks.
2026-09-25
Cantonese AI Capability 2026: Research and Models
Three studies (arXiv:2408.16756, HKCanto-Eval, Measuring HK MMLU), model performance by task type, three Cantonese text forms, common errors, availability paradox.
2026-09-25
Kimi K3 Review: Architecture, Data, Deployment Risk
Kimi K3 MoE specifications, the gap between Moonshot figures and Artificial Analysis, always-on thinking, pricing, thinking-history sensitivity and regulatory risk.
2026-09-25
Grok vs Claude 2026: Benchmarks and HK Access
16-benchmark comparison of Grok and Claude, Chatbot Arena ELO, agentic and coding strengths, pricing structure, and the availability gap for Hong Kong users.
2026-09-25
Mainland Firms in HK: AI Cross-Border Compliance
PIPL treats Hong Kong as overseas, DSL jurisdiction limits, the Greater Bay Area MOU, OpenAI and Anthropic access limits, and a PIPL plus PDPO architecture.
2026-09-25
Hong Kong AI Regulation 2026: Frameworks and Gaps
DPO generative AI guideline, HKMA and SFC circulars, IA sandbox programmes, PCPD framework, DoJ working group, three gaps against the EU AI Act, compliance strategy.
2026-09-25
HK 2026 Digital Subsidies: Open, Closed, Outlook
TVP closed to applications on 31 December 2024, DTSPP applications closed, BUD Fund Easy BUD limit raised to HK$150,000, enhanced DTSPP expected in H2 2026.
2026-09-25
HK PDPO and AI Compliance: Minimum Requirements
How DPP1, DPP3 and DPP4 apply to AI systems, the PCPD AI framework and RMC contract clauses, section 33 never in force, prescribed consent, and a minimum checklist.
2026-09-25
Hong Kong Education AI Policy: DSE, IELTS, CFA
EDB Circular 227/2024 and the AI science scheme, no standalone HKEAA AI policy, DSE English SBA guidance, IELTS and CFA positions, PDPO duties for student data.
2026-09-25
HK Construction AI Documents: BIM Mandate Roadmap
Development Bureau Circular 1/2025, the 2029 BIM submission target, three document tiers and AI automation feasibility, AP liability, mixed-language OCR accuracy.
2026-09-24
Hong Kong Family Office AI Research: SFC Boundaries
Type 4 and 9 licences, SFC circular Ref 24EC55, AI research tasks for family offices, MNPI input risk and the HKMA supplementary framework.
2026-09-24
Hong Kong Logistics Document AI: Accuracy Limits
Three tiers of Hong Kong logistics documents, Cap. 60 deadline, 87% HS code accuracy case, C&SD EMAN results, dual classification and OCR limits.
2026-09-24
Healthcare Admin AI in Hong Kong: Data Boundaries
Hospital Authority AIDA platform, A&E wait data, private hospital deployment, PDPO and patient consent, and three admin tasks suited to automation.
2026-09-24
Hong Kong Retail WhatsApp AI: Meta Rules, Handoff
WhatsApp Business API pricing timeline, Cantonese test scenarios, Meta AI disclosure rules, four human handoff triggers and BSP onboarding for retailers.
2026-09-24
Hong Kong Accounting Firm AI: What to Automate
HKICPA position, four automatable accounting workflows, audit judgement that needs a licensed person, and data security gaps between firm sizes.
2026-09-24
AI Recruitment Screening in Hong Kong: Law and Bias
Hong Kong discrimination law and AI résumé screening, University of Washington and HKU bias data, PDPO rules for candidate data, pre-deployment bias tests.
2026-09-24
HK Insurance AI Underwriting: Rules and Deployment
IA AI Cohort Programme, GenA.I. Sandbox++, Insurance Ordinance position, PCPD data guidance, mixed-language forms and anti-discrimination compliance.
2026-09-24
Hong Kong Law Firm AI Review: Limits and Liability
HKLS position paper, Judiciary guidelines, AI in contract review, liability, review records, LPP risk and two fabricated-citation cases.
2026-09-24
Business Email AI: Triage, Drafting, Follow-Up
Business email AI at three points: classification, drafting and follow-up reminders, with accuracy, human review ratio and the AI-human division of work.
2026-09-24
Meeting Minutes AI: From Recording to Action Items
Four stages of meeting minutes automation: transcription, transcript cleanup, LLM summaries and action item extraction, with a four-tool cost table.
2026-09-10
AI-Assisted Research: Hallucination Detection, Citation Verification and Source Tracking
Three hallucination types in AI research output: factual errors, fabricated citations, disconnected sources. Detection signals, a three-step verification workflow, and a framework for positioning AI in the research process.
2026-09-10
AI for Health Checkup Reports: Capability Limits and Three Safety Principles
Where AI adds real value — terminology, reference range context, preparing doctor questions. Where it cannot substitute — personal history, urgent signal combinations, clinical judgement. Three safety principles.
2026-09-10
AI-Written Cover Letters: Structural Advantage and Personalisation Limits
AI cover letters pass structural checks but consistently fall short on specificity. Three common misuse patterns, the correct three-step workflow, and why raw material quality determines AI output usability.
2026-09-10
Perplexity vs Claude: Fundamental Differences in Use Case Positioning
Perplexity for real-time retrieval with citations; Claude for reasoning synthesis and multi-step analysis. Different accuracy risk profiles and how a hybrid workflow outperforms either tool.
2026-09-10
Claude Memory vs ChatGPT Memory: How Architecture Shapes Long-Term Experience
AI-led automatic memory versus user-led explicit memory — how memory is created, controllability differences, scope limits, and the memory staleness problem both systems face.
2026-09-10
Claude Model Tier Selection: Haiku to Fable — Capability Boundaries and HKD Cost
Haiku, Sonnet, Opus, Fable — capability positioning, HKD cost estimates per million tokens, selection framework, and why multi-model routing outperforms a single-tier approach for enterprise systems.
2026-09-10
AI Memory Corruption: Identifying Three Error Types and How to Correct Them
Stale facts, over-generalised preferences, contaminated context — three memory error types with symptom patterns, identification signals, and why in-conversation correction alone is insufficient.
2026-09-10
Seed Memory Design: What to Include and What Causes Harm
Stable background facts, communication preferences, standing work context, specific constraints — four types worth including. Why temporary states and contradictory instructions cause output drift. Maintenance cadence guidance.
2026-09-10
Prompt Engineering in Practice: Five Fixes for AI Output Quality
Role framing, format declarations, step-by-step reasoning, negative examples, iterative refinement — five prompt design fixes with before-and-after examples for different output failure modes.
2026-09-10
Personalised AI vs Generic Chatbot: The Engineering Source of the Experience Gap
How memory retrieval logic, persistent extraction, and priority handling create the experience gap between stateless chatbots and personalised AI systems.
2026-09-10
Five Practical Business Uses of AI Conversation History
Decision audit, style learning, error tracking, knowledge accumulation, preference calibration — five uses each with different system design requirements. Treating conversation history as infrastructure is the key shift from tool to personalised assistant.
2026-09-10
Why AI Memory Cannot Cross Platforms: Architecture Limits and Migration Reality
Cloud-locked, project-scoped, and app-layer custom storage — three architectures. Semantic format differences, non-portable vector indexes, and divergent injection logic make cross-platform migration nearly impossible.
2026-09-10
AI Memory vs RAG: Two Architectures Solving Fundamentally Different Problems
Memory targets personalisation; RAG targets knowledge expansion — different sources, update methods, and serving scope. Priority design when both run simultaneously and the most common confusion between the two.
2026-09-10
Neon Postgres and pgvector: RAG Database Design for Production
pgvector IVFFlat vs HNSW index selection, Neon Serverless connection pooling, cold-start latency, and hybrid search design — four decisions cheapest to make at design time.
2026-09-10
Embedding Model Selection for Production RAG: Four Evaluation Dimensions
OpenAI text-embedding-3 vs Gemini text-embedding-004 vs open-source local models. Traditional Chinese mixed-language considerations and why embedding model names must be named constants.
2026-09-10
AI Application Log Design: What to Record, What Must Go to Database
Console.log in serverless is retained 24–72 hours only. Ten fields every LLM call must persist to database, time zone standardisation, and meaningful alert design.
2026-09-10
Multi-Model Routing: Dynamic LLM Selection by Task Complexity
Flagship model costs 10–30× lightweight models, but most tasks don't need it. Task classification dimensions, classifier design options, fallback chains, and principle-based routing rules.
2026-09-10
Silent Failure in AI Applications: Every Error Must Leave a Trace
Three high-frequency silent failure patterns: API timeout, rate limit, empty response. Minimum logging fields every LLM call must record and why explicit failure-path design is non-negotiable.
2026-09-10
When AI Context Window Overflows: Four Engineering Strategies
Truncation, summarisation, RAG replacement, pagination — four context management strategies with trade-offs. The real failure is deploying with no management mechanism at all.
2026-09-10
LLM Streaming in Production: SSE, Keepalive, and Reconnection Design
Three production streaming pillars: intermediate proxy timeout, server-side keepalive, client reconnection logic. Neglect any one and streaming breaks under long tasks or mobile networks.
2026-06-21
Why Enterprises Need an AI Consultant, Not Just Claude or Poe
Already using Claude, ChatGPT, or Poe? Six common functionality gaps and three layers of execution capability gaps, with a framework for choosing self-build vs. a consultant.
2026-07-09 · Tool
AI Automation Cost Calculator
Four numbers, instant monthly HKD comparison of manual vs AI processing. All calculation assumptions disclosed and adjustable.
2026-07-09
Why SME AI Projects Fail: Six Recurring Patterns
The problem is almost never the model. Six recurring structural failure patterns and how to prevent each.
2026-07-09
Fixed-Scope vs Time and Material: How to Price AI Automation Projects
Same project, two pricing models, completely different risk distribution. A contractor's first-hand analysis.
2026-07-09
Hire a Full-Time AI Engineer or Use a Contractor: A Decision Framework
Four decision dimensions and an often-overlooked hybrid model, with an honest case for both sides.
2026-07-09
The True Cost of AI Automation: Four Cost Layers Beyond the Development Quote
The development quote is only the first layer. Four layers to calculate before signing.
2026-07-09
How to Calculate AI Automation Costs: 2026 Method and LLM API Pricing for Hong Kong SMEs
Open formula, disclosed assumptions, and LLM API prices verified in July 2026. Calculate with your own numbers.
2026-07-09
ChatGPT API Pricing 2026: Comparison with Claude, Gemini and HKD Conversion
Subscriptions and API billing are two separate systems. July 2026 official prices from three providers, with HKD conversion.
2026-07-09
AI Document Processing Automation: Converting Unstructured Documents to Structured Data
Four workflow types, five system components, and the honest truth about accuracy rates.
2026-07-09
How to Evaluate an AI Vendor Proposal: A Non-Technical Buyer's Checklist
Zero technical background needed. Three steps: test with your data, three output dimensions, eight contract questions.
2026-07-09
Why Source Code Handover Matters: The Core Buyer Protection Clause
The most overlooked clause in AI contracts — and the one with the largest long-term cost impact.
2026-07-09
An AI Automation Project from Consultation to Delivery: The Complete Six-Stage Workflow
Six stages from requirements to source code handover — what happens, what to do, and what to receive.
2026-07-09
Five Questions to Answer Before AI Adoption
"Should we use AI?" is the wrong first question. Five better ones, answerable in one internal meeting.
2026-07-31
Prompt Caching in Production: How Enterprises Reduce LLM API Costs
Billing mechanics, 10% Cache Read rate, four silent failure modes including dynamic content and TTL format errors. Production data: long conversations save 80–90%.
2026-07-31
Gmail PDF Automation with Google Apps Script: No Server Required
Monitor Gmail, extract PDF attachments, call Gemini for AI summaries, push via Telegram. Real shipping company deployment, six-minute limit, and upgrade criteria.
2026-07-31
Three Layers of Enterprise Document Automation: From Format Conversion to Decision Support
Why most enterprises stop at layer one. Three questions before AI adoption, technical architecture flow, and a phased implementation framework.
2026-07-31
Agentic AI Is Losing Meaning: Three Questions That Cut Through the Vendor Pitch
Two proposals both labelled Agentic AI can be entirely different systems. Three questions any buyer can ask — no technical background required — to tell them apart.
2026-07-31
RAG Accuracy Depends on Document Type: Engineering Realities for Enterprise AI Q&A
List, paragraph, and semantic documents each require different chunking strategies. Vector search limits on open-ended queries and four conditions to verify before trusting any accuracy claim.
2026-07-09
Enterprise Data and AI: Privacy Questions Before Sending Documents to an AI System
Where does your data actually go? Plain-language explanation plus six privacy questions to ask before signing.
2026-06-25
Why AI Projects Fail in Hong Kong: Three Procurement Traps Only an Engineer Can See
It's not the model. It's not the budget. Confidence boundaries, production cost modelling, LLM fallback — three structural problems that don't appear in sales proposals but every engineering design review checks.
2026-06-25
What Is MCP? The New Infrastructure Standard for Enterprise AI Integration in Hong Kong
MCP is appearing in Hong Kong job postings and procurement requirements. For non-technical decision-makers: how this standard affects vendor dependency, expansion cost, and three questions to ask before signing.
2026-06-24
AI Consulting in Hong Kong: Six Questions Before You Buy
The vendor market is expanding fast and proposal language is converging. Six questions to ask before engaging an AI consultant — from process analysis capability to access control design.
2026-06-24
WhatsApp AI Integration in Hong Kong: Five Backend Questions Before Deployment
Meta policy narrows the frontend choice set. What determines system quality is knowledge management, hallucination handling, conversation storage, system integration, and platform dependency — five backend design decisions.
2026-06-24
Copilot, Gemini or Custom System? A Generative AI Decision Framework for Hong Kong Enterprises
88% of Hong Kong enterprise employees already use AI tools, but most usage stays at the individual level. The capability boundaries between Copilot, Gemini, and custom systems are determined by use case, not brand.
2026-06-19
Where Production AI Actually Lives: Why Traditional Enterprises Are the Better AI Consulting Bet
Traditional industries are AI consulting's real blue ocean. Logistics, insurance, finance, manufacturing — real problems, real budgets, missing only the people who can turn workflows into production AI systems.
2026-06-13
When People Leave, Where Does the Knowledge Go? Hong Kong's Most Expensive Hidden Cost in 2026
Staff turnover is not just a headcount problem. Business knowledge buried in emails and WhatsApp can now be made queryable — but architecture needs to be designed for it.
2026-06-13
After AI Delivery: Why Full Source Code Handover Is the Real Ownership Test
How dependency gets designed into AI systems, and how to identify it before you sign. Four verifiable standards for true ownership.
2026-06-13
Your Documents Are Half Chinese, Half English — This Is Where Most AI Systems Fail
Cantonese queries, Traditional Chinese documents, English clauses — why benchmark accuracy doesn't hold in Hong Kong. Three verification actions before you sign.
2026-06-13
"Does My Data Go to America?" — What Hong Kong Businesses Need to Know About AI Data Flows
Consumer ChatGPT and enterprise API are two different frameworks. Data flows, contractual protections, and architecture design — make an informed decision.
2026-06-07
AI Automation for Hong Kong SMEs: Build Once, Own It
HK$300M government funding is entering the market. Why a purpose-built AI system costs less and does more than a subscription stack for SMEs with a clear workflow.
2026-06-07
AI Agents vs Fixed Pipelines: How to Choose the Right Architecture
Every vendor proposal includes "intelligent agents." Why fixed pipelines win on cost predictability, auditability, and fault isolation — plus a single whiteboard test for your use case.
2026-06-07
When an LLM Gets It Wrong: Accountability and Verification in Production AI Systems
Hallucination is a statistical property, not a bug. Retrieval grounding, confidence boundaries, evals, audit logging — and four questions to ask any AI vendor before deploying in a regulated industry.
2026-06-06
Fine-Tuning, RAG, or Prompt Engineering? A Practical Decision Guide for Enterprise Buyers
Most business use cases don't need a custom-trained model. What each approach actually does, the real cost of fine-tuning, and three questions to ask before signing.
2026-06-06
Why You Can't Hire an AI Engineer in Hong Kong — and What the Alternative Looks Like
Large institutions want senior engineers, Web3 wants research skills, and the middle is empty. Why a full-time hire is the wrong structure for most SMEs, and what project-based engagement fixes.
2026-06-05
AI News Monitoring: The Data Source Problem Nobody Quotes You On
LLM summarisation is solved. The real question is what data can legally enter your system. A factual breakdown of public sources, paid subscriptions, and social media before you receive a quote.
2026-04-30
Most People Pick AI by Brand, Not Quality — Gemini's App Store #1 Is a Warning Sign
App Store rankings measure downloads, not accuracy. Gemini's hallucination problem and how to pick the right AI for your use case.
2026-04-30
Same Price, Very Different AI: What HK$155/Month Actually Buys You
ChatGPT Plus, Claude Pro, and Gemini Advanced cost the same — but hallucination rates, memory systems, and reasoning ability differ significantly.
2026-04-30
Your Family Member Has Cancer and You Can't Read the Report — How AI Can Help
Pathology reports, oncology terminology, treatment options — AI isn't your doctor, but it's your best translator. A practical guide for cancer caregivers.
2026-05-01
AI Engineer Hong Kong Freelance: What to Actually Expect in 2026
What production AI engineering actually costs, how long it takes, and how to avoid wasting time on the wrong hire.
2026-04-20
If AI Doesn't Know Who You Are, It's Just a Search Engine
Real AI memory has two layers: what you tell it, and what it learns from every conversation. Without both, it's just answering questions.
2026-04-28
Why AI Memory Is the Most Underrated Feature: Four Mechanisms Compared
Four memory mechanisms, Claude vs ChatGPT vs Gemini, setup guide, contamination risk — why memory is the most underrated AI feature.
2026-04-21
Upload a PDF to Claude and Read It 10x Faster
Contracts, reports, academic papers — practical guide with real prompts and one risk most people overlook.
2026-04-23
Why Does AI Know You Less the Longer You Use It? A Real Comparison of 3 Memory Systems
ChatGPT remembers you but gets the priorities wrong. Claude barely remembers at all. Memory without priority is not real understanding.
2026-04-26
Why Does Claude Feel Different Depending on Where You Use It?
The model hasn't changed — but the parameters might be very different. Context window, web search loops, output limits, memory: four variables most users never see.