Day 2 · R1 Industry Applications
Four speakers cover the industry-wide landscape of LLM applications, data governance in music streaming, retail recommendation and agent strategy, and 30 years of enterprise AI deployment experience.
At a glance
| Speaker | Talk | In one sentence |
|---|---|---|
| Chen Tun-ho (陳敦和), Product Consultant, IISI Group | Taxonomy of LLM Development and Applications in Industries | Uses two tracks — open source and proprietary — to map the full LLM ecosystem, and shares enterprise cases such as nursing shift handover and natural-language querying of Excel, SQL, and Cypher. |
| Kuan Shun-hui (官順暉), KKTech | Orchestrating the Future: AI Innovations in Music Streaming | Using KKBOX music data as a case study, discusses data quality governance, an LLM-powered playlist DJ, and an experiment that uses audio embeddings alone for "song recognition by ear." |
| Happy Lee (李昆謀), CPO, 91APP | Retail AI: from recommendation to agent | From traditional machine-learning clustering and OpenAI embedding-based fuzzy search, to a person/product/tag three-tower model and a future vision of a retail agent ecosystem. |
| Lin Hsien-cheng (林縣城), CTO, Innovation Research Center, GSS | Enterprise AI Application Practices – Expect the Unexpected! | Reviews the company's 30 years of AI investment from expert systems to LLMs, showcasing deployed products such as a chatbot, trademark image search, three-statement financial report recognition, and official document drafting. |
4 of 4 talks
1Taxonomy of LLM Development and Applications in IndustriesChen Tun-ho (陳敦和), Product Consultant, IISI Group (資拓宏宇); formerly with Taiwan AI Academy
Uses two main tracks — open source and proprietary — to map the full landscape of LLM technology, and shares multiple enterprise deployment cases.
Key points
- Opened with a self-introduction: currently a Product Consultant at IISI Group (資拓宏宇); previously worked at Taiwan AI Academy helping with industry training and adoption; before that, worked in the mobile phone industry.
- Divides the LLM development process into three stages — pre-train → fine-tune → RLHF: fine-tuning is further split into full fine-tune, continued pre-training (CPT, which sits between full fine-tune and pre-train, used for the "TAIDE" model), and partial fine-tune (e.g., adapters, LoRA).
- Mentioned knowledge distillation: using another model (e.g., ChatGPT) to generate Q&A data to train one's own base model, noting that OpenAI has stated this practice violates its terms of use; Alpaca and Vicuna are academic research projects of this type built on LLaMA.
- Divides proprietary-model applications into four categories: Communication (prompt engineering, likened to "communicating with a body of knowledge"), Enhancement (multimodal input such as images/video, plugins that supplement model capabilities such as math, RAG that supplements knowledge), Interoperability (external/internal APIs), and Customization (e.g., ChatGPT's GPTs).
- Shared the "Nursing Assistant" product: uses an LLM to classify patients' health issues, combines it with RAG to retrieve nursing data, and automatically generates shift-handover records, significantly saving nurses' time.
- Demonstrated the concept of "tool operation": using natural language to operate Excel (generating formulas for customer case analysis), relational databases (natural language to SQL for querying carbon inventory reports), and the graph database NEO4J (natural language to Cypher syntax to build a person-relationship graph).
- Proposed the view that "chemical molecules, code, robot control, and time-series values can all be treated as a form of language": using SMILES (Simplified Molecular Input Line Entry System) to represent chemical molecules as an example, noting that SenseTime used a similar language-model concept for drug molecule generation even before ChatGPT appeared; the same logic extends to using LLMs for "predictive generative AI" on factory time-series data.
- Shared a food-delivery satisfaction evaluation case: treats structured ratings (food temperature, packaging, service attitude, etc.) and open-ended text feedback both as "language" input, training a model to generate customer-service reply text.
- Company overview: IISI Group (資拓宏宇) is a major Taiwanese ICT company that has long helped the government build transportation, financial, and healthcare systems, and offers an integrated deep AI product suite (chatbot, knowledge management, content generation/summarization/conversion modules).
Tech, products & figures
- LLaMA, Mixtral
- Open-source base models (base model / foundation model)
- LoRA (Low-Rank Adaptation)
- A lightweight fine-tuning technique that adjusts only a small number of added parameters. Original paper: Hu et al., 2021, "LoRA: Low-Rank Adaptation of Large Language Models" (arXiv:2106.09685); official implementation: microsoft/LoRA (GitHub).
- Chain of Thought (CoT)
- A prompting technique; the speaker likened it to "explaining things step by step when communicating with a person."
- RLHF (Reinforcement Learning from Human Feedback)
- Makes model output better aligned with human value judgments
- Alpaca, Vicuna
- Academic models built on LLaMA, trained via knowledge distillation from ChatGPT. Alpaca official pages: Stanford CRFM Alpaca blog, stanford_alpaca (GitHub); Vicuna official page: LMSYS Org Vicuna blog.
- NEO4J + Cypher
- A graph database and its query language, demonstrated by using natural language to generate Cypher and build a person-relationship graph. Official documentation: Neo4j Cypher Manual.
- SMILES
- A simplified text notation for chemical molecules, used by SenseTime in drug-molecule generation research. Background: Simplified Molecular Input Line Entry System (Wikipedia); original paper: Weininger, 1988, "SMILES, a chemical language and information system."
Notable quotes
A large language model is a body of knowledge — although it has no life, it's actually just like an ordinary person: it's about how you communicate with it.
Q&A
- No Q&A session (the moderator moved directly to the next speaker).
Fact-check notes
- The speaker's company, IISI Group (資拓宏宇), is a Taiwan-based ICT service provider, stock code 6614.Sources：iisigroup.com資拓宏宇 Facebook
- The speaker's claim of having worked at Taiwan AI Academy is consistent with public records of "Chen Tun-ho / AIA Deputy Director of Technology Division," but the public-source basis for the "Product Consultant" title at IISI Group could not be further verified (unverified).Sources：台灣人工智慧學校 2020 meetup 頁面
2Orchestrating the Future: AI Innovations in Music StreamingKuan Shun-hui (官順暉), KKTech (科科科技); introduced by the moderator as "Executive Vice President," listed in the official agenda as Chief Scientist (see verification notes)
Using KKBOX's music data as a case study, emphasizes "govern data quality first, then talk about AI," and demonstrates a music-search experiment based purely on audio embeddings.
Key points
- The moderator introduced the speaker's expertise as streaming technology, recommendation systems, and computer graphics, describing him as an important founding contributor to KKTech's technology development; the speaker joked that "today is probably the only talk that's actually about music."
- Opening figures: ten years ago KKBOX added about 800 songs per day (about 300,000 per year); last year (2023) that had grown to about 250,000 songs per day (nearly 100 million per year), highlighting that the scale now far exceeds the number of songs a person could listen to in a lifetime.
- Breaks down "the data of a song" into five categories: the music itself (waveform), lyrics (the textual extension of music), metadata (song title, composer, performer, genre, publisher, etc.), commercial/copyright data, and user behavior data.
- Case one (metadata quality): using the video game soundtrack for Ghost of Tsushima as an example, notes that metadata errors caused other works by the same composer to be unsearchable on KKBOX; further analysis found genre classification to be chaotic — sampling showed "Pop" and its variants (e.g., Mandarin Pop) were over-concentrated, the same anime genre was split into multiple different genre tags, and Brazilian country music (Sertanejo, which the speaker described as the era's Latin-American equivalent of a "City Pop"-level hit genre) had 12 different spellings in the database.
- Identifies the sources of metadata problems as including typos, duplicates, uneven distribution, aliasing (alias synonyms), and outliers (e.g., songs with a year field of 190 BCE or 9999 CE).
- The team first benchmarked against other platforms (e.g., KKBOX has over 30,000 genre tags vs. Spotify's roughly 6,000) to confirm this really was their own data anomaly rather than "nitpicking," then used Multilingual BERT + HDBSCAN to cluster genres and re-consolidate genre classification (also noting that Spotify and the ISMIR academic community are tackling the same problem).
- Case two (interaction/recommendation): treats the LLM as "your own DJ," letting users describe a situation in natural language (e.g., "I'm feeling really down," "I want company on Saturday afternoon") to get a playlist; the approach converts data into embeddings for similarity retrieval (fuzzy search) plus content filtering (to avoid offensive or inappropriate content); but also stresses that when user intent is already clear (e.g., "I want 90s rock"), direct keyword search (old technology) is more accurate and cheaper than forcing the use of LLM/embeddings.
- Case three (pure-audio embedding search, a proof of concept completed in 6 weeks with two interns from National Chengchi University): using no text/metadata at all, only the song's audio itself (waveform → embedding, with the model referred to by the speaker as "mast," see verification notes), demonstrates "find similar songs by listening," including: using a high-energy segment of Mayday's "倔強" (Stubborn) to find several other Mayday songs; using Cheer Chen's "太聰明" (Too Smart) with vocals removed, searching on the instrumental alone and still finding instrumentals with a similar feel; testing the system with an "AI Stefanie Sun" cover of a Jay Chou song, where the system honestly failed to find any of Stefanie Sun's actual works (because the catalog contains no original Stefanie Sun recordings) — it was instead people outside the system who were fooled into thinking it sounded like her; using "Beautiful in White" for extended wedding-playlist recommendations.
Tech, products & figures
- Multilingual BERT, HDBSCAN
- Used for semantic clustering of genre tags, consolidating duplicate/similar genre classifications
- librosa
- A Python audio-processing library, used for energy-based segment extraction
- The "mast" model
- The name of the model the speaker used to generate audio embeddings; its exact full name could not be verified (unverified) (possibly a music-audio representation model similar to MERT / MusicFM, but this cannot be confirmed).
- Cosine similarity
- Used for similarity comparison in the embedding vector space
Notable quotes
If your data is kind of not okay, then you can expect everything is not okay at all.
The "AI Stefanie Sun" is terrible... if anyone gets fooled, that's a human problem, not a machine problem.
Q&A
- No Q&A session (the moderator moved directly to the next speaker).
Fact-check notes
- The official agenda lists the speaker's title as "KKTech Chief Scientist," but the moderator introduced him as "KKTech's Executive Vice President." A search also found that a "Kuan Shun-hui (Drake)" holds the public title of "Senior Technical Director" at KKStream, a subsidiary of KKCompany (formerly CTO of KKStream). The three sources (official agenda, moderator's introduction, external report) disagree on the title, and it could not be confirmed whether all three refer to the same person at the same point in time, so all three versions are presented as-is without determining which is authoritative (unverified).Sources：官方議程頁 conf2024.aiacademy.tw/agendaKKStream 案例報導，AI Foundry Edge
- KKBOX is a music streaming service under KKTech (KKCompany).Sources：數位時代報導kkcompany.com
- The name of the "mast" audio-embedding model remains marked unverified: no corresponding public model documentation or paper could be found, and it cannot be ruled out that this is an abbreviation of a different model name (similar technologies such as MERT or MusicFM are only an inference by this note, not a verification result).
- Internal figures such as KKBOX's daily song uploads (about 800 ten years ago, about 250,000 last year) and genre tag counts (KKBOX's 30,000+ vs. Spotify's roughly 6,000) are all the speaker's own account, with no external corroboration.
3Retail AI: from recommendation to agentHappy Lee (李昆謀), CPO, 91APP
Moves from traditional machine-learning clustering to embedding-based fuzzy search, arriving at a "person/product/tag" three-tower model for retail and a future vision of multi-agent collaboration.
Key points
- The moderator introduced the speaker as Chief Product Officer of 91APP; the talk focuses on how retail can use AI to solve the fundamental matching problem of "people finding products, products finding the right people."
- Recalls that the company invested in traditional machine learning as early as 5–6 years ago (around 2018–2019): using 20–30 features such as purchase count, browse count, and category tags to cluster products; and using clickstream patterns to analyze intent patterns such as users "just browsing," "interested in a specific product," or "trying to reach the free-shipping threshold," producing a purchase-intent clustering model called "DCIU" — the high-purchase-intent segment is only about 1% of members, but after being reached by marketing campaigns it can contribute about 50% of overall campaign revenue.
- After generative AI emerged, tried using OpenAI's embedding model (ADA; the speaker stated the dimensionality verbally as "1,356," while OpenAI's official documentation confirms 1,536 dimensions, see the technology field below) to create product embeddings, using cosine similarity to find similar products, and once published an e-commerce recommendation plugin on the ChatGPT Plugin Store (since discontinued, later evolved into GPTs).
- Using pure text-embedding fuzzy matching for products ran into problems: for example, when a user specified a "white" product, the matched results were semantically close but the color didn't match (because retail product descriptions often use marketing terms like "pearl white, snow white, matte white, dream white," which traditional word segmentation can't catch, but the embedding space semantically clusters these terms); the approach was later changed to first filter by keywords/tags and then perform embedding-based tag-expansion search, which markedly improved accuracy.
- Demonstrated a ChatGPT Plugin demo launched in 2023: given the input "Mid-Autumn Festival gift, not a traditional mooncake, something creative," the system recommended drip-coffee bags, pineapple cakes, and handmade soap shaped like pastries, among others.
- Argues that the real challenge in retail is that "consumers often can't articulate what they want" (they buy what they "want," not what they "need"), so external tags and behavior must be collected to build a persona; developed the "Two-Tower Model" (the shared architecture behind YouTube/Spotify/Netflix recommendation systems) into three vector spaces — person, product, tag — that retrieve against each other (six combinations, such as person-to-product, product-to-tag, tag-to-person, and so on).
- Application scenarios: personalized ("a thousand faces for a thousand people") recommendation systems, automatic product tagging (also referencing Google SEO optimization), and ad audience packages (person → tags → similar potential audiences → likely products as dynamic creative); also notes that different placements have different goals (homepage emphasizes CTR, cart page emphasizes conversion rate), requiring a re-ranking layer after vector retrieval that sorts by objective (traditional machine learning, not GenAI).
- Internally, the company named the whole vector-retrieval-plus-LLM-spokesperson architecture "Joy" (a pun, since the pronunciation is close to "91"), which can be composed into agents for customer-service Q&A, product recommendation, and more.
- Envisions future retail scenarios: agents will appear on both the consumer and seller sides (e.g., a smart speaker placing orders for consumers), potentially leading to "agent-to-agent" transactions (likened to wholesale-market bargaining that outsiders can't follow); retail organizations will move toward headless (service API-ification) to accommodate the many channels and agents connecting in.
Tech, products & figures
- OpenAI embedding model ADA
- The speaker stated the dimensionality verbally as 1,356; OpenAI's official documentation confirms 1,536 dimensions. Official documentation: OpenAI, "New and improved embedding model"; OpenAI Vector embeddings guide; text-embedding-ada-002 model page.
- ChatGPT Plugin Store
- A third-party plugin store that existed briefly in 2023, later evolving into GPTs
- Two-Tower Model
- A common architecture in recommendation systems such as YouTube, Spotify, and Netflix; here extended into three towers — person, product, tag. The technical origin is generally traced to the YouTube recommendation paper: Covington et al., 2016, "Deep Neural Networks for YouTube Recommendations" (ACM RecSys); architecture explanation also in Google Cloud Blog: "Scaling deep retrieval with TensorFlow Two Towers."
- Retrieval + Re-rank (RR architecture)
- First performs vector retrieval, then re-ranks according to business objectives (CTR / conversion rate)
Notable quotes
What consumers actually buy isn't what they need — it's what they want.
You can't just sell them what they need — you have to sell them the need they don't even know they have.
Q&A
- No Q&A session (the moderator moved directly to the next speaker).
Fact-check notes
- The official agenda and public records confirm the speaker is Li Kun-mou (李昆謀), English name Happy Lee, currently Chief Product Officer (CPO) of 91APP.Sources：91APP 產品長專訪，關鍵評論網LinkedInhappylee.blog
- The speaker stated OpenAI's ADA embedding dimensionality verbally as 1,356 dimensions; OpenAI's official documentation (see the technology field link above) confirms 1,536 dimensions; this note was verified against the official documentation, not a fabricated figure.
- The DCIU purchase-intent model's figures — "the high-intent segment is about 1% of the base, contributing about 50% of campaign revenue" — are the speaker's own account of internal company data, with no external corroboration.
4Enterprise AI Application Practices – Expect the Unexpected!Lin Hsien-cheng (林縣城), CTO, Innovation Research Center, GSS (叡揚資訊)
Reviews the company's 30-year AI journey from expert systems to LLMs, showcasing multiple enterprise application products already deployed.
Key points
- Opens by paying tribute to the Transformer paper "Attention Is All You Need," likening the company's philosophy to "Application Is All We Need."
- Personal career retrospective: his first job in 1990 was already in knowledge-base systems / inference engines (the expert-system era), working on expert systems such as China Steel boiler fault diagnosis and disease diagnosis; later went through an AI winter due to the "knowledge acquisition and representation" bottleneck; then went through Data Mining (recommendation engines, once led a team to the US for R&D), neural networks / genetic algorithms, deep learning, and up to generative AI in recent years.
- Describes the common feeling among enterprises adopting AI as "hopeful yet afraid of getting hurt": hopeful that new technology can do what was previously impossible, yet afraid that resources just invested will be made obsolete by the next new technology (the anxiety of "money down the drain"); argues there is no one-size-fits-all technology, and the right method should be chosen based on the problem itself, resources, and requirements.
- Over the past decade the company has invested in four major technology areas: text recognition (NLP), computer vision (image recognition), chatbots, and, in the past year or two, LLM applications; organizationally it adopts a "central kitchen" model of technology R&D that then empowers product lines such as official document systems, knowledge management, and net-zero carbon emissions.
- Enterprise AI adoption is divided into five stages: proof of concept (POC) → real-world implementation → process optimization → becoming part of the company's DNA → generating a competitive impact.
- Demonstrated multiple product demos:
- Chatbot platform: applied in hospital scenarios, and converting sales staff's voice/text call reports into knowledge-management platform assets.
- Trademark image-search retrieval system: already live in production use at the Intellectual Property Office, with a feedback loop (manual annotation of similar/not similar) that continuously retrains to improve accuracy.
- Handwritten triplicate form / invoice recognition: printed-text recognition accuracy is close to 100%; handwritten Chinese recognition accuracy is lower but still practically useful; high-speed rail ticket recognition is nearly 100%.
- Automatic recognition of the three financial statements (balance sheet, income statement, etc.): used in SME loan review workflows, with over 90% accuracy; customers report saving more than 80% of manual data-entry time, with suspected anomalous figures flagged in red (combined with post-processing business rules, such as subtotal-sum verification).
- Introducing LLMs into the knowledge-management platform: emphasizes that in RAG scenarios, "retrieval" quality matters more than "generation" quality (the speaker estimates that roughly four-fifths of the effort is in preprocessing and retrieval rather than generation itself); the quality of multimodal document extraction (including charts) directly affects downstream retrieval usability.
- Official document system: the earliest deployment was Chinese grammar/spelling correction (spelling/grammar check, whose technology has evolved through multiple generations up to BERT), stressing that general-purpose models (e.g., using ChatGPT directly) are not necessarily adequate given enterprise-specific terminology and higher error-tolerance requirements; the current system can help recommend document templates and generate drafts of official documents/meeting notices.
- Personal observations on future trends (listed but not elaborated due to time constraints): the open source vs. proprietary debate will continue; enterprise customized models don't necessarily need to be large models; multimodal documents will become a key application; knowledge graphs (which the company invested in 5 years ago) combined with LLMs will produce more precise applications for complex scenarios; agent development; MLOps/LLMOps engineering needs; data quality and quantity remain the biggest bottleneck; compute resource constraints; AI governance and evaluation standards (how to objectively prove one's own model is better than others'); and continued evolution of model architectures (Transformer is currently best but not necessarily the final answer).
Tech, products & figures
- BERT
- Used for Chinese grammar correction and entity recognition (NER) in official documents; the speaker noted that not every task requires an LLM
- RAG (Retrieval-Augmented Generation)
- The primary application pattern for enterprise knowledge management; the speaker emphasized that front-end retrieval/document-extraction quality is the key bottleneck
- Knowledge Graph
- The company invested in this 5 years ago; combined with LLMs for more complex scenario applications
- Three-statement financial report recognition, trademark image search, handwritten triplicate-form OCR, official document system
- All are actual deployed products/customer cases of the company
Notable quotes
We're both hopeful and afraid of getting hurt. Hopeful that something new comes along that lets us do what we couldn't do before; but afraid of getting hurt in the sense that after I've invested, a new thing comes out a few days later, and then wasn't my investment money down the drain?
Attention is all you need — but in fact, for GSS, or for this industry, what we'd really say is Application is all we need.
Q&A
- No Q&A session (the program moved directly into the next session after the speaker finished).
Fact-check notes
- The speaker's company, GSS (叡揚資訊), has stock code 6752.Sources：gss.com.tw公司簡介頁
- A search found Lin Hsien-cheng's current title to be "R&D Director / Deputy General Manager of GSS," which does not fully match the "CTO, Innovation Research Center" title listed in the official agenda; both are nonetheless senior technical-management roles at GSS, and this is judged to be a title difference for the same person across different points in time / different sources.Sources：叡揚大事紀
- The trademark image search, three-statement financial report recognition, and official document system demos shown by the speaker were all customer cases / internal product presentation screens; a check of GSS's official website found it only publicly lists its Vital-series cloud services (e.g., Vital CRM, Vital Finance, Vital BizForm, Vital Knowledge), with no public product page found that directly corresponds to what the speaker showed. Therefore the details of the trademark-search, financial-statement-recognition, and document-drafting demos (recognition rates, deployment scope, etc.) are all the speaker's own account, with no external corroboration.Sources：公司簡介頁