By the numbers
Every figure quoted across the two days, gathered in one place — useful when you need to put a number in a proposal. Each row names the speaker or organisation it came from.
Talks noted40 talks
Speakers45 +
Sourced references90 +
Parallel tracks3 rooms
Deployment outcomes
| Domain | Speaker / organisation | Outcome |
|---|---|---|
| Financial fraud | E.SUN Financial | Alert-account detection cut headcount 50%; about 4,000 accounts flagged 2023–2024, preventing over NT$70m of further loss |
| Card fraud | E.SUN Financial | About NT$120m of loss prevented per year |
| E-commerce listing | 91App | Rule + AI + human mix, throughput up 10× |
| Factory issue triage | Pegatron | Agent accuracy 97.7% (only five simple tools attached) |
| Defect root cause | Pegatron | 94% accuracy on an IC substrate case |
| Warehouse ABC analysis | Kenmec | 2 months → 2–3 days |
| Line commissioning | Merck Kaohsiung (Siemens case) | Dry-run time cut by up to 50% |
| Brewery energy | Heineken (Siemens case) | 20% energy saved and 50% CO2 cut per plant, rolled out to 15+ sites |
| Grid transmission | German customer (Siemens case) | Dynamic Line Rating lifted throughput on the same line 20%, with no extra hardware |
| Meeting minutes | AIA (IRB case) | 72 person-hours → about 2 |
| Travel customer service | Lion Travel (agent Lily) | 37,449 calls Jan–Jul, 13% at night; about NT$231,700 of revenue in the last two months |
| Travel labour | Lion Travel smart lockers | About NT$4m a year in labour overhead saved |
| Insurance marketing | iCook Intelligence (Japanese P&C group) | Revenue up 60%, renewal rate up 80%; launch time from 1–2 years to 1–2 days |
| Sports injury prevention | NTU of Sport (Pingjhen High School trial) | Zero pitcher injuries over the trial academic year |
Technical figures
| Topic | Source | Figure |
|---|---|---|
| On-prem vs cloud break-even | Cisco | At 250 inference requests per minute, about 1.5–2 years to break even against cloud token cost |
| AI inference traffic growth | Cisco internal | About 40% monthly, 46× annually; each agentic engine averages 8–14 repeat confirmation calls |
| Vertical vs general model | APMIC | 24B vertical beats 120B general on Taiwanese law; VRAM 50GB against 292GB |
| GRPO training gain | Twinkle AI | F1 model on MMLU: 50 → 64 |
| Agent accuracy vs MCP tool count | Salesforce, cited by Pegatron | Down to about 33% in some scenarios; into the teens with open models |
| Markerless 3D joint reconstruction | NTU of Sport | About 3 cm in 2024 → about 2 cm in 2025 |
| Multi-sensor online calibration | Tron Future Technology | 2 hours → 15 minutes; angular error 45° → under 1° |
| IMU-only drift with no GPS | Tron Future Technology | Hundreds of metres of error within about 40 seconds (exponential divergence); pure visual relocalisation about 20 m |
| Smart badminton venue | NYCU | One RTX 4090 handles 8 cameras in real time (measured ceiling about 12) |
| Anomalous-sound recognition | Deep Wave | Human ear about 80%, machine about 90–95% in a controlled environment — but not on a real production floor, so it was never productised |
| Police drone response | US, cited by the drone association | 71 seconds average to scene; Phoenix PD flies 10,000+ missions a year |
| Bridge crack detection | Taiwan Hope Innovation | A photograph every 50 cm at 0.2 mm resolution |
| AI clam grader | STUST | 600 clams a minute per line against 2,800 by hand — hence the plan for ten lines |
Costs
| Item | Source | Cost |
|---|---|---|
| Humanoid robot, per unit | Kenmec | About NT$4m; training 20 could reach hundreds of millions |
| Force plate | NTU of Sport | 60×60cm about NT$600k; a 10-metre plate about NT$10m |
| UWB positioning system | NTHU | About NT$3m per set |
| Venue hire and setup | NTHU | About NT$200k a game, just to install and remove the floor logo |
| AR glasses | STUST demo | Pre-production about NT$200k; mass-produced about NT$30k; Xiaomi AI glasses about RMB 1,999 |
| Inertial navigation component | Taiwan Hope Innovation | About NT$300k years ago (import required a signed undertaking) → about US$3 today |
| Image generation | Accu Crazy | Nano Banana about NT$10 an image, against about NT$500k for a traditional shoot |
| Multi-agent newsletter | AIA class example | About NT$6.6 per run |