sciwork seminar 2026

To see our schedule with full functionality, like timezone conversion and personal scheduling, please enable JavaScript and go here.
09:00
09:00
30min
Registration

Registration

non-talk
R122
09:30
09:30
10min
Opening

Opening

non-talk
R122
09:40
09:40
30min
Reports on Collaborating with Claude and Codex on Large Projects and Challenging Tasks
En Shih

Can Claude be considered AGI? Over the past three months, I having testing the limits of how well Claude can do on the most challenging tasks that I encounter. This is a first-hand report of my sessions, crystallized from burning over 40 billion tokens, and spanning across topics including RF, FPGA, GPS, PCB, 3d rendering, and trading.

「跟 Claude 和 Codex 一起搞大型專案和硬核任務的實戰心得」
Claude 算得上 AGI 嗎?這三個月來,我一直在拿手邊最棘手的任務狂操 Claude,看看它到底能做到什麼程度。這篇是我親身上陣的第一手紀錄,燒掉超過 400 億 tokens 換來的心得,內容橫跨 RF、FPGA、GPS、PCB、3D 渲染、還有交易等各種領域。

Technical Talk
R122
09:40
30min
Understanding FP8 Numerical Spaces in LLM Quantization
Yu-Fang Hu

FP8 is becoming increasingly important in modern LLM systems, but its behavior is often difficult to reason about intuitively. Unlike integer quantization, FP8 combines floating-point’s non-uniform numerical spacing with quantization scaling challenges. In this talk, we will explore FP8 formats, outlier problems, and scaling effects through visualization, with the goal of building stronger intuition for low-precision floating-point quantization in LLMs.

Technical Talk
R115
09:40
120min
現代流程引擎工作坊:簡化跨服務、多步驟流程開發 Part 1
Jiang

你是否曾感到痛苦:「跨服務任務一多,重試邏輯與狀態轉移就寫到崩潰?」 讓我們停止自行維護脆弱的流程,利用現代化流程引擎,將模糊的步驟轉化為強韌的自動化工作流!

Workshop
R330
09:40
120min
突破 SQL 查詢極限:用 Graph 思維處理複雜關聯與多跳推論 Part 1
Laurence Chen

很多工程團隊都曾遇過這類問題:

JOIN 越寫越長,query 幾乎無法維護
「找關聯」的需求開始出現多跳 traversal
recommendation、權限繼承、dependency、fraud detection 等問題很難用 table 直觀表達
application layer 出現了「查資料 → 組關係 → 再計算」的邏輯
這類問題的共同點是:

問題的核心其實是「關係」,但我們仍然被迫用表格思維處理。

本 workshop 將透過 Graph Analysis 的方式,重新理解這類問題。

Workshop
R329
10:25
10:25
30min
Building AI Agents in the Real World - 從 Prompt Engineering 到 Harness Engineering 的 AI Agent 工程實戰
KK

LLM 很強,但真正困難的從來不是「叫 AI 回答問題」,而是如何讓 AI 在真實世界穩定工作。
這場分享將從 AI Agent 的工程實戰出發,討論如何從單純 Prompt Engineering,逐步走向 Context Engineering 與 Harness Engineering,建立真正能在產品、工作流程、企業場景中運作的 AI Agent 系統。

Technical Talk
R122
10:25
30min
VMPilot: A Modern C++ Virtual Machine Protection Framework
scc

VMPilot is an advanced virtual machine protection framework implemented in C++. Secure by design, VMPilot is specifically engineered to safeguard your software from reverse engineering. Offering seamless integration and ease of use for your projects, VMPilot sets a new standard for software protection.

Unlike traditional black box solutions, VMPilot is built with transparency in mind. Its inner workings are easily understandable, yet formidable to crack. By incorporating modern cryptography and obfuscation techniques, your software is shielded against potential attacks. Even with the computing power of a supercomputer, breaking VMPilot in parallel becomes a daunting challenge.

Technical Talk
R115
11:10
11:10
30min
From Optimal Control to Reinforcement Learning for Quadrotor Drones
Sheng-Wen Cheng

This talk introduces the basic ideas behind quadrotor optimal control and reinforcement learning from a practical, beginner-friendly perspective. We will discuss how position, velocity, attitude, and angular velocity are used to describe the drone state, and how a controller converts tracking errors into motor commands.

Several representative control approaches will be introduced, including the Linear Quadratic Regulator (LQR), H∞ robust control, Model Predictive Control (MPC), and Reinforcement Learning (RL). The talk will explain the intuition behind each method: what it tries to optimize, what kinds of problems it handles well, and what trade-offs appear in practice.

The examples will be based on RotorBench, an open-source Python-based quadrotor control benchmarking framework. More details about the project are available at: https://github.com/shengwen-tw/rotor-bench

Technical Talk
R115
11:10
30min
Mini Taiwan!把台灣開放資料,交給 Agent 養成一套會自己長大的系統
Migu

Mini Taiwan 系列,是想要將跟台灣有關的空間開放資料,讓他們在地圖中真的運作起來。

這場分享會拆解一個資料分析師如何與 AI 協作,把台灣的開放資料做成「Mini Taiwan」這套會自己長大的系統。內容聚焦在「怎麼把它做成系統」這件事:從幾個已上線的作品出發,談我與 Claude Code 的分工,接著是系統的組成——資料收集、資料庫與 schema、資料生命週期,以及如何用同一套骨幹長出不同的應用。最後分享目前在自動化上的嘗試,包含一個把驗收納進每一步的開發循環。

Technical Talk
R122
11:40
11:40
80min
Lunch

Lunch

non-talk
R122
13:00
13:00
30min
FP4: A P4-based Network Controller for Load Balancing
c1ydeh

The data plane needs low latency to support high-speed packet forwarding. In this talk, I will start with an interesting load balancing problem, then propose a controller-based solution using a P4-programmable switch. Finally, I will demonstrate how the controller changes the network behavior after the solution is applied.

Technical Talk
R115
13:00
30min
在 2026 年淺談 Web DevOps
TerryChan

2025/2026 年 ai agents百花齊放, 用 ai agents 打造出 web apps 非常容易
我想淺談中小型流量的系統, 從後端設計、開發、測試到部署與可觀測性
有哪些坑容易踩到, 需要注意哪些細節, 把我蒸餾成 skill 分享給大家

Technical Talk
R122
13:00
120min
現代流程引擎工作坊:簡化跨服務、多步驟流程開發 Part 2
Jiang

你是否曾感到痛苦:「跨服務任務一多,重試邏輯與狀態轉移就寫到崩潰?」 讓我們停止自行維護脆弱的流程,利用現代化流程引擎,將模糊的步驟轉化為強韌的自動化工作流!

Workshop
R330
13:00
120min
突破 SQL 查詢極限:用 Graph 思維處理複雜關聯與多跳推論 Part 2
Laurence Chen

很多工程團隊都曾遇過這類問題:

JOIN 越寫越長,query 幾乎無法維護
「找關聯」的需求開始出現多跳 traversal
recommendation、權限繼承、dependency、fraud detection 等問題很難用 table 直觀表達
application layer 出現了「查資料 → 組關係 → 再計算」的邏輯
這類問題的共同點是:

問題的核心其實是「關係」,但我們仍然被迫用表格思維處理。

本 workshop 將透過 Graph Analysis 的方式,重新理解這類問題。

Workshop
R329
13:40
13:40
80min
Coding Sprint

modmesh / commitizen-tools

Sprint
R122
13:40
80min
Coding Sprint

swportal / BBGo

Sprint
R115
15:00
15:00
60min
Refreshment

Refreshment

non-talk
R122
16:00
16:00
50min
Coding Sprint

modmesh / commitizen-tools

Sprint
R122
16:00
50min
Coding Sprint

swportal / BBGo

Sprint
R115
16:50
16:50
10min
Closing

Closing

non-talk
R122