Day 2 · R3 Deep Tech Flash Talks
Nine flash talks spanning marketing AI, LLM security attacks, temple-tech applications, no-code automation for non-engineers, multimodal smart manufacturing, eval-driven development, and AI-powered energy savings — showcasing the diversity of deep-tech deployment.
At a glance
| Speaker | Talk | In one sentence |
|---|---|---|
| Wu Wei-han (吳威翰), Founder & CEO, Accucrazy (肖準) | Craft your Growth Strategy with a Fully Fine-tuned AI Brain | Using its own marketing data to fine-tune the Moana model, mass-producing word-of-mouth posts that pass the Dcard Turing test. |
| Shen Yi-ting (沈宜婷), SOC / Security Engineer, CHT Security (中華資安國際) | Insecure llama?: When LLM becomes a weapon for hackers | Using a locally deployed, restriction-removed Ollama model to automate penetration testing, successfully breaching a Linux system to obtain root. |
| Li Ming-da (李明達), Founder, MYAI168.COM / LEADERG (立達軟體科技) | The Application of the Latest AI Algorithms in Temple Field | Using LLM vision and a fine-tuned large model for temple fire/theft prevention and a fortune-slip interpretation robot, with a live demo on stage. |
| Li Mu-yue (李慕約), Curator, Generative AI Conference | How non-engineers can use AI to write code and leave work five minutes earlier | Teaching non-engineers to use ChatGPT to generate Google Apps Script for automating expense reports, surveys, and computer-use-based online ordering. |
| Lin Yu-hung (林育弘), Director, AI System Planning Division, Contrel Technology (東捷科技) | Multimodal AI Empowering Smart Manufacturing Upgradation | A panel/semiconductor laser-repair equipment maker integrating imagery, measurement data, and generative AI for multimodal defect detection and repair. |
| Kao Chi-an (高季安), General Manager, Foresight Technology (先知科技) | AI-Driven Manufacturing Innovation: Visual AI Platforms and Generative Image Technology | A no-code image defect detection platform using generative data augmentation to solve manufacturing's high-mix, low-volume pain point of insufficient samples. |
| Chang Wen-tien / iHoward (張文鈿), Founder, iHower Information (愛好資訊) | Eval-Driven-Development | Proposing a 5-level maturity model for LLM application development, demonstrating AI-driven automatic prompt generation and optimization. |
| Hu Hsiang-wei (胡翔崴), Engineer, ITRI | Advancing Medical Innovation with Multimodal Edge Computing | A multimodal medical large model linking compounds, physiological signals, and imagery for health-education Q&A, standardized patient training, and 3D dental model generation. |
| Hsieh Tsung-chen (謝宗震), CEO, Chimes AI (詠鋐智能) | Decarbonizing Energy-intensive Industry: How AI Solutions Accelerate ESG Goals | Using predictive plus generative AI to help factory equipment save energy — a case study cut annual electricity use by 32%, with over 40,000 tons of cumulative carbon reduction. |
9 of 9 talks
1Craft your Growth Strategy with a Fully Fine-tuned AI BrainWu Wei-han (吳威翰), Founder & CEO, Accucrazy (肖準行銷集團) marketing group; its product line includes The Pocket Company
Fine-tuning the Moana model on the company's own marketing data to write word-of-mouth posts that fool Dcard users.
Key points
- The speaker previously served as Global Vice President of Meitu and as a TikTok Business Development lead, then moved into the metaverse/Web3 space before fully pivoting to AI starting in 2023.
- The product The Pocket Company is positioned as a company you can run from your pocket, with compute partners AWS, NVIDIA, and Taiwan Web Service (台智雲).
- Observed the evolution of marketing trends: TV ads, then major KOLs (such as Tsai A-ga), then group-buying/livestream commerce, then Threads — where the key is authenticity and personalization rather than needing to be an influencer.
- Fine-tuned the first through fourth generations of the Moana model on large volumes of the company's own historical marketing data; it can automatically generate word-of-mouth posts to match a persona (such as an "auntie" or "little sister") and a product angle, using emoji and parenthetical asides to sound like an authentic Taiwanese internet-forum voice.
- Case study: an AI-generated post on Dcard drew over 1,300 comments with no one identifying it as AI-written, and it was even featured by Dcard's own "Shui Shui Lab" curated section; feeding the same prompt to ChatGPT produced writing that felt noticeably more like sponsored content.
- Business model: a sponsored post that normally costs NT$30,000–40,000 on the open market costs only about NT$60 to produce with AI.
- Argued that fine-tuning captures small, delicate, tacit knowledge (such as a specific cultural tone of voice) better than RAG or LoRA; RAG, by nature, still relies on the base model's original capability with knowledge merely bolted on externally.
- Envisioned a multi-agent architecture: one agent defines the target audience, Moana writes the post, an agent called Coco schedules posting through virtual KOC accounts, and an agent called Rebecca converts the content to voice for release — forming a fully automated pocket-sized marketing team.
Tech, products & figures
- The Pocket Company
- An AI marketing platform under Accucrazy (肖準行銷).
- Moana
- A word-of-mouth post generation model fine-tuned on the company's own marketing data (1st through 4th generation).
- Dcard Shui Shui Lab
- Dcard's own official curated content section.
Notable quotes
"We found that the AI produced a kind of language no human had ever spoken, yet it still sounds understandable to us."
"This is actually a real Turing test... not a single one of those comments suspected we were AI."
Q&A
- No formal Q&A; after the talk, the moderator briefly remarked that these four agents are just like your own employees.
Fact-check notes
- Accucrazy/Xiaozhun Marketing, The Pocket Company, and Moana were all verified as accurate; Wu Wei-han is indeed the founder and CEO, and previously served as Global Vice President of Meitu.Sources：Accucrazy 官網TechNice 專訪
- The speaker title "Founder & CEO" on the official agenda matches the verification results.
2Insecure llama?: When LLM becomes a weapon for hackersShen Yi-ting (沈宜婷), SOC / Security Engineer, CHT Security (中華資安國際)
Using a locally deployed, restriction-removed Ollama model for automated penetration testing that successfully breached a Linux system and obtained root.
Key points
- Speaker background: has published multiple security research papers and discovered vulnerabilities in Google and some open-source platforms, focusing mainly on connected-vehicle and LLM security; this talk credits the HackingBuddyGPT team's research framework as its extension base.
- Framing: most "hacking GPT"-style projects use OpenAI's cloud models as the attack engine; this talk instead used a locally deployed Ollama model to test how far an attacker could get with a small, local model.
- Most models (the speaker estimated over 80%) have built-in safety restrictions and will refuse to produce malicious commands; however, unrestricted variant models exist (the live transcription rendered the name phonetically as something close to "Bei-la-ma," likely some kind of uncensored/abliterated model) that can be told to do anything and will bypass safety restrictions.
- Compared human red-teamers/penetration testers with LLM-automated attackers: humans rely on deep, pre-existing attack skills and experience-based judgment, while LLMs rely on built-in automated attack scripts and still need humans to supply some contextual information.
- Demo flow: the target was set as a Linux host, with the LLM generating attack commands. It initially issued random commands and meaningless symbols, showing it hadn't truly understood the task; after the script was modified, the LLM successfully breached the system with a normal-privilege account, found the root password (password1234), and obtained root access.
- After obtaining root, the LLM kept automatically trying different commands to dig up more sensitive information (such as /etc/passwd, the shadow file, and password hash files), logging the entire attack process.
- Speaker's conclusion: industry discussion of LLM security today mostly stays at the theoretical level (basic protections like access control), and there needs to be more proof-of-concept-level empirical research, rather than just talking about restrictions and regulations.
Tech, products & figures
- Ollama
- A local LLM deployment tool, and the vehicle for this talk's attack experiment.
- HackingBuddyGPT
- The open-source hacker-GPT research framework the speaker credited as a reference.
- Prompt Injection
- Mentioned as one of the attack techniques used to make the model bypass its safety restrictions.
Notable quotes
"Today we're playing the bad hacker, not the good hacker."
"Everyone really needs to build attack defenses for these LLM models... you need much more detailed testing, proofs of concept, and so on, to test whether an attack can actually succeed."
Fact-check notes
- Verification: LinkedIn shows that Shen Yi-ting works at CHT Security (中華資安國際).Sources：沈宜婷 LinkedIn
- The term "Bei-la-ma" in the live transcription is unclear in meaning; it is presumed to refer to some kind of uncensored model, but the specific corresponding model name could not be confirmed (unverified).
- Related LLM security discussion also appears in notes from other sessions (Yang Cheng-lin / CyCraft on privacy, alignment, and localization risks of LLM adoption and RCE vulnerability research in RAG frameworks; Weng Hao-cheng on the evolution of GenAI security attacks; Cha Shih-chao on AI security governance).Sources：R1-產業應用.mdR3-深度技術.md
3The Application of the Latest AI Algorithms in Temple FieldLi Ming-da (李明達), Founder, MYAI168.COM / LEADERG (立達軟體科技); Adjunct Associate Professor at National Taiwan University and National Taiwan University of Science and Technology
Building temple fire/theft-prevention systems and a fortune-slip interpretation robot with LLM vision and a fine-tuned large model, with a live fortune-slip demo on stage.
Key points
- LEADERG Software Technology (立達軟體科技, stock code 7505, capitalized at NT$200 million) mainly sells AI computers, software, firmware cards, and solutions; its official site is myai168.com.
- AI Door God System: uses a large language model's vision capability to detect small wisps of smoke or fire, replacing traditional smoke detectors (which have high false-alarm rates in a temple's incense-burning environment) and YOLO-type algorithms (which frequently misjudge small smoke/fire); its advantages are that it can be applied without extra training and has high accuracy. It can also detect suspected theft of incense-money donations, and upon detection it alerts temple staff via a Windows-side voice alarm plus a LINE group push notification with an on-site photo.
- AI Foot-Traffic Counting System: installed at temple entrances/exits, claimed to count up to 300,000 visits per day with zero error, and can record video playback for review.
- AI Building Exterior Inspection System: uses drones to periodically photograph temple carvings, combined with a large language model's vision function to identify possible damage or areas needing repair after earthquakes or typhoons.
- AI Image Recognition for Fuji (Planchette Writing) Archiving: recognizes and archives the often-illegible handwriting produced during spirit-writing (fuji) sessions, claiming to be able to process 30 years and over 100,000 spirit-writing documents at a self-reported 100% accuracy; the recommended hardware is an RTX 4090 24G, with a full setup costing roughly NT$200,000.
- AI Customer Service / Fortune-Slip and Scripture Interpretation: uses models of 70B parameters or larger (the speaker said testing showed models below 70B lack sufficient comprehension), using in-context learning plus full-parameter fine-tuning; explicitly does not recommend RAG (prone to retrieval errors causing inaccuracy) or LoRA (degrades the model's base capability). It can interpret the Wealth God Sutra, answer questions like "how can I get rich," and translate classical-Chinese fortune-slip poems into vernacular language with guidance. Setting up a fortune-slip-interpretation service is recommended to use four RTX 6000 Ada cards (a full setup costs about NT$1.8 million, with add-on features costing roughly another NT$400,000).
- Noted that large models, by default, tend to refuse to answer serious questions (such as about marriage, health, or company bankruptcy), and require fine-tuning before they will respond directly.
- Live demo on stage: the moderator and two audience members drew fortune slips live (asking about career, a family member's health, and whether a husband was having an affair, respectively), and the system instantly output plain-language interpretations and advice (such as "proceed cautiously, don't rush for quick results," "adjust your mindset and accept the advice," and "stay calm and communicate rationally"); the service is already live.
Tech, products & figures
- AI Door God System, AI Foot-Traffic Counting System, AI Building Exterior Inspection System, AI Fuji Archiving Recognition
- LEADERG's series of products built for temple scenarios.
- in-context learning / full parameter fine-tuning
- The technical approach the speaker argued is superior to RAG/LoRA (this is the speaker's own view, not a general consensus).
- Recommended hardware
- RTX 4090 24G (about NT$200,000) / 4x RTX 6000 Ada (about NT$1.8 million).
Notable quotes
"This is really the key point — worshippers feel that a plain-language answer matters a lot to them, because otherwise they draw a fortune slip and just can't understand it."
Q&A
- The session itself doubled as a live-demo interaction: the moderator and two audience members drew fortune slips and asked questions on the spot (career / a family member's health / suspicion of a husband's affair), and the system instantly output interpretation results (see the key points above).
Fact-check notes
- LEADERG Software Technology (立達軟體科技, stock code 7505), the MYAI168.COM education platform, and Li Ming-da as founder were all verified as accurate.Sources：MYAI168.COM台灣公司網
- Figures such as "100% fortune-interpretation accuracy" and "zero error" are all the speaker's own claims with no third-party verification found, and should be treated as vendor marketing language.
- Li Ming-da's stance against RAG/LoRA and in favor of full-parameter fine-tuning, together with Tsai Tsung-han (training a small model outperforms a general large model plus RAG) and Yang Li-wei (enterprise knowledge management should switch to RAG with swappable models), forms a third position in this year's conference's fine-tuning-vs-RAG debate; a summary of the three-way clash appears in a separate key-points summary document, section 4.Sources：R0-下午-智慧轉型與紀念講座.md重點摘要.md
4How non-engineers can use AI to write code and leave work five minutes earlierLi Mu-yue (李慕約), Curator, Generative AI Conference
Teaching non-engineers to use ChatGPT to generate Google Apps Script for automating expense reports, surveys, and computer-use-based online shopping.
Key points
- Core argument: one of ChatGPT's major potentials is translating Chinese into code, letting people who can't program still build automation, which lowers the barrier to coding.
- Demo 1: fed ChatGPT an Excel file with multiple sheets (customer data plus birthdays) and simply asked it to merge the sheets (equivalent to a VLOOKUP); ChatGPT used its built-in code-execution environment (Code Interpreter) to complete the merge and even proactively suggested calculating ages.
- Demo 2: used ChatGPT to generate Google Apps Script code, pasted it into the Google Apps Script site to run, and within seconds got a link that created a Google Form (for a post-event survey), skipping the need to manually set up each option.
- Demo 3 (expense-report automation): when the accounting contact changed and demanded a whole year's taxi receipts be reformatted, the speaker instead used ChatGPT-generated Apps Script: first labeling Uber-related Gmail messages, then using Apps Script to save the labeled emails as PDFs to Google Drive, and finally using another generated script to parse amounts and dates from all PDFs in the folder and compile them into a Google Sheet. He also shared a tip (credited to James Huang) for auto-pulling a trip's purpose from the calendar to avoid manually filling in expense-report descriptions.
- Recommended free resource: a website called the "Sense AI Conference Gift Pack," which breaks tutorials down into difficulty-graded steps such as one-star (doable in seconds) and two-star (takes a few minutes).
- Case study: a Sense AI meetup once invited a 9-year-old speaker to share how they used AI over an entire summer, iterating through daily conversations, to build an Apple Watch app — the emphasis being not on being a "child prodigy" but on the process of continuous, iterative dialogue with AI.
- Other resources: a free public talk by Amazing Talker's HR/customer-service/marketing managers on using AI to write Apps Script for internal automation; an "internet hotline" Q&A board; a "case library" collecting real-world cases of AI writing code; a monthly book club; and an online course (with a discount code for NT$249 to catch up on the previous day's workshop content).
- Closing demo (computer-use / AI agent): gave the instruction "go to PC Home 24hr, buy a case of apple juice, pay with stored-value balance, don't ask me about the process, just place the order"; the AI agent used computer vision to read the screen and autonomously selected the product, entered the quantity, checked out, and paid with the stored-value balance, successfully placing the order — the physical goods were eventually delivered to the speaker's home.
Tech, products & figures
- Google Apps Script
- A programmable "little computer" behind Google's services that can operate Gmail, Drive, Sheets, Calendar, and more.
- ChatGPT's built-in code-execution environment (Code Interpreter)
- Used to directly process uploaded Excel files and documents.
- Computer-use-style AI agent
- Uses computer vision to read the screen and simulates mouse/keyboard actions to complete an online purchase.
Notable quotes
"AI is a 20-dollar magic wand — a lot of people may actually have magic in them, but before, we didn't have this wand. Now that you have this wand, you can do some pretty amazing things."
"Don't ask me about the process, just do it... all I need to do is drink the apple juice."
Q&A
- No formal Q&A; after the talk, the moderator praised the "don't ask me about the process, just place the order" segment.
Fact-check notes
- Li Mu-yue is indeed the co-founder and curator of the Generative AI Conference.Sources：Podcast 節目介紹
- "Sense AI" is presumed to be a self-created brand name close to the speaker's community "Sun*" or something similar; the exact correct name could not be verified.
5Multimodal AI Empowering Smart Manufacturing UpgradationLin Yu-hung (林育弘), Director, AI System Planning Division, Contrel Technology (東捷科技)
A panel/semiconductor laser-repair equipment maker integrating imagery, measurement data, and generative AI to build a multimodal defect detection and repair system.
Key points
- Contrel Technology's (東捷科技) core product is laser-repair equipment, with customers mainly in the panel and semiconductor industries; its subsidiary Holytech (和來科技) operates automation equipment for the logistics industry (such as vision-based package sorting).
- In recent years the company has actively invested in AI systems integration, having placed third in the Ministry of Economic Affairs-hosted IDEAS Show commercial image-AI recognition competition, and has an industry-academia collaboration with Professor Lien Cheng-chieh's team at National Cheng Kung University.
- Definition of "multimodal AI": it's not just about linking together different data forms (text/image/speech) — the key is that different single-modality models can exchange information with each other without needing a human to manually re-integrate each model's output.
- Illustrated the efficiency advantage of multimodality: a compound query that used to require three separate steps — a language AI, a vision AI (reverse image search), and a conversational AI — (such as "find this restaurant's best dish and read out its name in Italian") can be handled by a multimodal model in a single pass.
- Real-world application: from upstream AI inspection equipment (scanning defect locations) and measurement equipment (such as whether line width is out of spec) to downstream laser-repair equipment, multimodal data is linked together for semantic understanding, making repair paths more precise and potentially even replacing some manual repair work.
- Inventory of single-modality technologies: image AI focuses on defect/anomaly detection and labeling (using anomalization-type models); generative AI is used for super-resolution synthesis of low-resolution images to compensate for industrial camera hardware limits; data-type AI mostly uses AutoML-integrated models to diagnose abnormal values; reconstruction AI uses DFD technology to generate height (depth) information from 2D images; and path-planning AI plans repair paths by defect type, aiming to match the results of manual hand repair.
- Outlook: once multimodality matures, smart factories will become more efficient, more flexible, and better able to support small-batch customized production; human-AI collaboration models will vary by industry (in some cases AI executes and people coordinate, in others AI assists while people execute); improved efficiency in upstream data collection will also drive industry upgrading, echoing the Industry 4.0 concept.
Tech, products & figures
- Anomalization model
- Used for image anomaly/defect detection.
- DFD technology
- Used to generate height/depth information from 2D images (reconstruction AI).
- AutoML
- Used for anomalous-value diagnosis in data-type AI.
- IDEAS Show commercial image-AI recognition competition
- Hosted by the Ministry of Economic Affairs; a team from Contrel Technology's subsidiary once placed third.
Notable quotes
"Models on their own actually can't share support with each other, but multimodality makes that possible... they're able to interoperate with one another."
Q&A
- No formal Q&A; the moderator's closing remarks briefly echoed the impact of multimodality on smart manufacturing.
Fact-check notes
- An online search could not find public records of Lin Yu-hung working at Contrel Technology's AI System Planning Division, but the company's background (laser-repair equipment, panel/semiconductor customers, and its subsidiary Holytech doing logistics) closely matches the content of the on-site talk.Sources：東捷科技官網
- The speaker's title "Director of the AI System Planning Division" is consistent with "Director" on the official agenda.
6AI-Driven Manufacturing Innovation: Visual AI Platforms and Generative Image TechnologyKao Chi-an (高季安), General Manager, Foresight Technology (先知科技)
A no-code image defect detection platform that uses generative image augmentation to solve manufacturing's high-mix, low-volume pain point of insufficient samples.
Key points
- Speaker background: over 30 years of experience in equipment, process engineering, and information technology; collaborated with Professor Cheng Fan-tien of National Cheng Kung University on an AI project during his doctoral research, then founded Foresight Technology (先知科技) to bring TSMC-style smart-manufacturing concepts to small and medium enterprises.
- Foresight Technology's business covers generative AI, RPA + AI, image-application AI, and data AI, and can build hybrid multimodal applications with clients (such as combining engineering-drawing text with image analysis to improve profit margins).
- Identified three major manufacturing pain points: (1) manufacturing sites are scattered, so systems need cloud-based, cross-site capability; (2) clients generally lack AI or even IT talent (joking that Taiwan's talent mostly goes to startups or big companies); (3) it's hard to build discriminative/decision-type AI with enough training data for high-mix, low-volume specialty products, since large orders tend to flow to the scale economies across the strait.
- Core product "TianRanTongHao" (天然通好): a no-code image defect classification/localization platform, claimed to handle everything from data collection, labeling, training, and evaluation to deployment in one stop; it automates hyperparameter tuning and model selection, and continuously embeds the latest algorithms the company has won in international competitions into the platform.
- Generative image-augmentation cases: in a beekeeping case, using just one real image paired with a large volume of generative deep-synthesis images for training achieved a validation-detection mAP of 0.785 (the speaker said 0.5 or above is the threshold for production-readiness); in a fastener bending-defect case, using only 2-3 real samples paired with generatively synthesized defect samples achieved a detection rate of nearly 100%.
- Conclusion: generative AI isn't only for generating content — it can also work together with discriminative/decision-type AI, breaking the conventional assumption that decision-type AI must have large amounts of real samples, which is especially suited to Taiwanese SMEs' high-mix, low-volume production style.
- Anecdote: shared a story of a client that built its own AI team to write prompts, where results from an ordinary internal user using the Foresight platform ended up only slightly better than what the company's own AI experts produced with hand-written prompts; the speaker reframed this as a positive case of "your employees have all become AI experts."
Tech, products & figures
- TianRanTongHao (天然通好)
- Foresight Technology's no-code image defect detection/classification platform.
- Generative image data augmentation (few-shot + synthetic data)
- Used to solve the problem of insufficient training samples in manufacturing's high-mix, low-volume production.
- mAP (mean Average Precision)
- An object-detection evaluation metric the speaker cited, reaching 0.785 in the case study.
Notable quotes
"Generative AI isn't only usable for generation — it can work together with decision-type AI to break through many baselines within decision-type AI that we used to think were impossible to break."
Fact-check notes
- Foresight Technology, Kao Chi-an as General Manager/founder, his doctoral studies at National Cheng Kung University's Institute of Manufacturing Information and Systems (in collaboration with Professor Cheng Fan-tien), and his prior employment at TSMC were all verified as accurate.Sources：先知科技官網報導INSIDE 專題
7Eval-Driven-DevelopmentChang Wen-tien (張文鈿), Founder, iHower Information (愛好資訊); both he and the moderator referred to him on stage by his English name "iHoward"
Proposing a 5-level maturity model for LLM application development, demonstrating AI-driven automatic test-data synthesis and prompt optimization.
Key points
- The speaker positions himself as an AI engineer, runs the "AI Engineer" newsletter, is well versed in OpenAI, Claude, and RAG technology, and personally maintains a Traditional Chinese embedding-model benchmark (which he says ranks first on Google search); his full slides are published on his personal blog, ihower.tw.
- Core argument: the answer to the uncertainty of LLM application (prompt) performance is to "do evaluation" — prepare a dataset, actually run it, and evaluate based on the output (which can be algorithmic evaluation, human evaluation, or AI scoring).
- Proposed a self-created 5-level prompt-engineering maturity model (compared against traditional deterministic software development): Level 0 — write the prompt with no testing or evaluation at all; Level 1 — just eyeball it once in the Playground; Level 2 — prepare a test-case set and manually analyze results to make adjustments; Level 3 — automated evaluation (analogous to traditional unit testing, but using an LLM or an algorithm to score automatically); Level 4 — analogous to TDD, achieving automated prompt optimization.
- Demonstrated "MetaPrompt" at Level 2: a prompt-generates-prompt tool already open-sourced by Anthropic (Claude), where entering a short task description generates a complete prompt (including few-shot examples); however, since the tool favors the XML-tag format Claude typically uses, switching to other models (such as OpenAI's) still requires manual format tweaking.
- Demonstrated "evaluation with a ground-truth answer" at Level 3: an initial Chinese-language prompt classification task achieved 87.5% accuracy, which rose to 95% after using MetaPrompt to generate an English-language version of the prompt; adding chain-of-thought ("think step by step") did not further improve accuracy, because the task itself was simple enough that the model didn't need extra reasoning steps to answer — this demonstrated how to use an objective score to decide how a prompt should be changed.
- Also demonstrated "evaluation without a ground-truth answer" at Level 3: writing an evaluation prompt that plays the role of a judge, listing 0-5 scoring criteria (LLM-as-judge); noted that human correction of AI scoring results is still needed, with corrected cases saved as few-shot examples fed back into the evaluation prompt, solving the "who watches the judge" problem. In a RAG context, relevance can be evaluated pairwise across question, retrieved data, and answer.
- Demonstrated automated prompt-optimization frameworks at Level 4: DSPy and TextGrad, both of which let you set which model synthesizes training data, which model generates the prompt, and which model makes the final prediction; after running for a few minutes, they automatically iterate to produce hundreds of candidate prompts and select the best version. In the demo case, the auto-optimized prompt for a classification task pushed accuracy to 100% (citing prior research, for example, that on the Llama 2 model, the best prompt for a math reasoning problem turned out to be a Star-Wars-style narrative, illustrating how inefficient hand-written prompts can be).
- Shared Colab notebooks for DSPy and TextGrad for the audience to take away and use directly.
Tech, products & figures
- MetaPrompt
- A prompt-generating tool already open-sourced by Anthropic (Claude).
- LLM-as-judge
- An automated evaluation approach where an AI scores outputs.
- DSPy, TextGrad
- Two prompt auto-optimization frameworks, both able to iteratively generate and select the best prompt.
- 5-level prompt-engineering maturity model (Level 0-4)
- A classification framework created by the speaker.
Notable quotes
"The efficiency of hand-writing prompts is very low."
Fact-check notes
- Chang Wen-tien (online handle ihower) is indeed the founder of iHower Information (愛好資訊, founded in 2018), specializing in the OpenAI API, Claude, and RAG, and runs the AI Engineer newsletter; being called "iHoward" on stage by himself and the moderator is likely a homophone/stage-name play on "愛好" (àihào), matching the official agenda's company name "愛好資訊"; the talk title "Eval-Driven-Development" matches the official agenda exactly.Sources：ihower.tw 簡報張文鈿 Facebook
8Advancing Medical Innovation with Multimodal Edge ComputingHu Hsiang-wei (胡翔崴), Engineer, ITRI (工業技術研究院)
A multimodal medical large model linking compounds, physiological signals, and imaging data for health-education Q&A, standardized patient training, and 3D dental model generation.
Key points
- Speaker background: has startup experience (co-founder of Xiang'an Biomedical and HIT for Asia), is a Phoenix Startup mentor at National Cheng Kung University, has worked deeply in AI technology for about 10 years, and does innovation R&D and publishes papers in both the finance and medical fields; he has already given nearly 20 talks this year.
- Multimodal technical pipeline: inputs such as text and images each pass through their own encoder, are then projected into a shared multi-dimensional vector space, and are mostly finally projected into an LLM, letting the LLM act as the multimodal bridge. In the medical field, the speaker extends this concept to "sequence-type data" such as compounds (SMILES strings), protein sequences, physiological signals (EEG, ECG), medical images, and medical-record reports — any sequence-type data can go through the same encoder-to-projection architecture.
- Application case 1 (doctor-patient Q&A / health education): an innovative approach published at the IQA Conference — having a second large model act as a "reviewer," scoring and giving feedback on the empathy and medical-consultation quality of the first model's response, then asking the original model to revise its response; the quantified result was a 14.29% improvement in professionalism and a 5.81% improvement in intent recognition.
- Application case 2 (personalized health education): built separate dietary-restriction databases for different stages of chronic kidney disease (such as stages 1 through 4), ensuring the system can give different, more precise personalized advice for the same question (such as dietary restrictions) based on the patient's actual stage, rather than a one-size-fits-all answer.
- Application case 3 (clinical report generation): in collaboration with a hospital, reads clinical data to automatically generate personalized care recommendation plans (covering cognitive training, light-and-music stimulation therapy, etc.); since clinical staff generally don't have time to hand-write personalized health-education sheets, this system scored well in clinical evaluation. The team also built a dementia-detection platform integrating vital-sign measurements such as eye-movement recognition, automatically producing charts and reports and supporting follow-up doctor-patient Q&A.
- Application case 4 (companion robot): built a companion robot on top of GPT-4o, adding a module for twelve eye-based emotional expressions and back-end health analysis; it has already been deployed in the long-term-care centers of two hospitals and was demonstrated on TPBS. It can chat, ask about everyday life, and also answer professional health-education questions.
- Application case 5 (standardized patient education system): reversed the traditional "virtual human educates the patient" model — instead, AI plays a "standardized patient" with a pre-set medical history (diabetes, hypertension, osteoporosis, etc.), letting trainee physicians practice taking a patient history; the system can automatically compare a physician's questioning results against the correct answers and give an automated score. About 8 hospitals are currently in ongoing collaboration talks, with a voice/virtual-human interaction interface planned to be added by year-end.
- Application case 6 (3D dental model generation): using fine-tuning (with a fine-tuned vs. non-fine-tuned comparison) together with the NVIDIA NeMo framework, first achieved text-and-image generation for medical-record reports, then demonstrated using text alone to directly control/generate a 3D tooth model (such as specifying a symptom like "outward-flared teeth"), moving toward automated generation from medical-record text to 3D imagery.
- Technical infrastructure: uses the NVIDIA edge-computing architecture (Jetson Orin series) and the NVIDIA NeMo framework, in ongoing collaboration with NVIDIA, with further experience-sharing planned for late October.
Tech, products & figures
- GPT-4o
- Cited by the speaker as a leading current multimodal model (low latency, strong voice generation and visual recognition).
- NVIDIA Jetson Orin series + NVIDIA NeMo framework
- The edge-computing infrastructure for the medical multimodal model.
- SMILES strings
- A serialized text representation of compounds, used to bring compound data into the multimodal encoder.
- IQA Conference
- The conference where the speaker published the paper on the "AI reviewer" mechanism.
Fact-check notes
- Hu Hsiang-wei's role as co-founder of Xiang'an Biomedical and HIT for Asia, and his R&D background at ITRI, were verified as accurate.Sources：遠見雜誌專訪
- Several result figures in the live transcription (such as "placed in the top 20 for GIN STAR" and "top 8 for LAMA", where the competition names were unclear in pronunciation) could not be verified (unverified).
9Decarbonizing Energy-intensive Industry: How AI Solutions Accelerate ESG GoalsHsieh Tsung-chen (謝宗震), CEO, Chimes AI (詠鋐智能); the moderator referred to him by his English name "Johnson"
Using predictive plus generative AI to help energy-intensive factory equipment save power — one case cut annual electricity use by 32%, with the platform's cumulative carbon reduction exceeding 40,000 tons.
Key points
- Speaker background: holds a PhD in Statistics from National Tsing Hua University, known as "the father of R in Taiwan," ranked second in Taiwan and 709th globally in GitHub contributions; leads the D4SG Data Heroes program, helping Taiwan use technology and data to transform society; he was a speaker at the very first edition of this conference's predecessor, the "Data Science Enthusiasts Conference," back in 2014, and is being invited again this year exactly 10 years later.
- Founded Chimes AI (詠鋐智能) with the goal of using AI technology to reduce waste and help organizations and companies accelerate ESG and carbon-reduction efforts; background argument: 78% of multinational brands require suppliers to provide a concrete timeline on energy conservation and carbon reduction.
- Case 1 (electric heating furnace): at a chemical plant in Kaohsiung, an electric heating furnace heats raw material to 300°C for over 30 hours per batch, consuming 20,000 kWh per batch (equivalent to one day's electricity use for 2,000 households), running about 60 batches a year for 1.2 GWh of annual electricity use and 600 tons of carbon emissions (requiring 50,000 trees to offset). Energy-saving operating advice from predictive AI plus generative AI (adjusting the heating curve and hold time while keeping the product within spec) achieved a measured 32% annual electricity savings, cutting annual carbon emissions from 600 tons down to roughly 200 tons (a reduction of 400 tons).
- Case 2 (semiconductor fab chilled-water system): a semiconductor fab's chilled-water system often accounts for 20-30% of the entire plant's electricity use; by fixing the required refrigeration capacity and adjusting the chilled-water system's tunable parameters, it's possible to save 5-15% of the chilled-water system's annual electricity use and extend equipment lifespan.
- Case 3 (equipment anomaly early warning): in partnership with Harbor Technology (皓博科技), vibration, current, temperature, and flow-rate data from equipment such as pumps, compressors, gearboxes, fans, blowers, mixers, and blenders are converted into an "equipment lifespan indicator" score for early warning; a generative AI assistant then explains possible failure causes and recommended actions, so inspection staff can be prepared before heading out, reducing labor time costs and industrial-safety risk.
- Scale and results: the "Tookie" platform (AutoML plus generative AI) has already been deployed in 50 factories with 2,500 AI applications, cumulatively saving over 40,000 tons of CO2 emissions per year.
- Product architecture: equipment sensor data is aggregated into a data middle platform, then fed into an AI PC running Tookie, which provides quality pre-diagnosis, real-time diagnosis, process operation optimization (saving electricity and carbon), and equipment pre-diagnosis ahead of unplanned downtime, using generative AI to translate predictions into recommendations that people can understand and trust.
Tech, products & figures
- Chimes AI (詠鋐智能)
- The company founded by the speaker, focused on AI-driven energy savings and carbon reduction.
- Tookie platform
- An integrated AutoML plus generative AI platform, used for equipment energy savings and anomaly early warning.
- Harbor Technology (皓博科技)
- A partner providing equipment vibration/current/temperature/flow-rate data.
Notable quotes
"There is only one Earth."
Q&A
- No formal Q&A; this was the last speaker of the entire R3 session, and immediately after, the moderator announced the end of R3 and directed the audience to the Closing Keynote.
Fact-check notes
- Hsieh Tsung-chen (Johnson Hsieh) is indeed the founder and CEO of Chimes AI (詠鋐智能), holding a PhD in Statistics from National Tsing Hua University, with the company founded in December 2020. The official agenda writes the company name as "詠鋝智能"; verification shows the correct form should be "詠鋐智能" (鋐, not 鋝) — this is a typo in the official agenda page itself, and this entry follows the verified form "詠鋐智能."Sources：未來城市＠天下專訪Chimes AI 官網