Chapter I · Problem / Product
April 2026

The worker is the last unmeasured variable in your factory.

  • On the same line, same shift, output between best and worst differs 20–40%.
  • Machines, materials, energy are all metered. The worker is not.
  • 20–40% of real work time is recorded by no MES, no WMS, no andon.
20–40%
OUTPUT GAP · SAME LINE · SAME SHIFT
ILO World Statistics · 2024 · n = 1,247 factories
Fig. 1.1 iFactory Watch — Specification2026-Q2 · v1.0 · BMI270 + nRF54L15
42 mm38 mmFRONT VIEW12.4 mm + sensorPPGheart-rateSIDE PROFILE01 STRAPsilicone, 22mm02 GLASSgorilla, 0.6mmOLED03 DISPLAY1.4" monoIMUMCU04 PCB · IMU · MCUBMI270 + nRF54L15180 mAh05 BATTERYLiPo, 5-6 day mixed-load06 BACK + PPGskin contact07 STRAPsilicone, 22mmEXPLODED VIEW · 7 LAYERSSAMPLING 100 Hz · 6-axis IMUBATTERY 180 mAh · 5-6 dayBOM $6.30 @ 10KRADIO BLE 5.2WEIGHT 32 g
VALIDATED
I
Chapter II · Deployment
WATCH $6.30 · SETUP < 1 day
II / one wristband · one gateway · one dock

One wristband, one gateway, one charging dock. Live in a day.

One gateway aggregates both streams via 5G/WiFi. Every worker visible within 30 seconds.

Fig. 2.1 Standard Deployment Topology — Factory Floor56 nodes · 1 gateway · 1 uplink
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 mDOCKOUTPUTGWRK35880102030405060708L10910111213141516L21718192021222324L33334353637383940L54142434445464748L64950515253545556L7↑ UPLINK · 5G/WiFifactory floor · ~1,000 m² · 56 stations · 1 gateway · 1 uplink to cloud engineACTIVEIDLEGATEWAY
Watch · every worker
SpecValue
BOM / worker$6.30 @ 10K
Sampling100 Hz IMU · 50 Hz PPG
RadioBLE 5.4
Battery180 mAh · 5–6 day
Visualization< 30 s

Raw IMU and PPG signals stay on-device. The gateway uploads only anonymized, aggregated behavioral features; raw data stays in China.

II · DEPLOYMENTII / VIII
Chapter III · Model
ARI 0.007 · AUC 0.832
III / Behavioral Model

Data in. Decisions out. Every recommendation attached to expected yield and ¥ output impact.

We do not sell dashboards. The supervisor sees concrete actions: who rests, who fills in, who needs retraining.

  • · Behavior clustering: ARI = 0.007 (12–39x tighter than self-report surveys).
  • · Bad-week prediction lift: AUC 0.832 vs 0.688 intensity baseline.
Fig. 3.1 Tangerine Behavioral Model — Architecturedata in · decisions out
INPUTS · 𝕏factory dataIMU100Hz · 6-axisℝ⁶⁰⁰CYCLEper-stationℝ⁶⁰DEFECTQC code{0,1}⁵⁰ROSTERshift · skillℝ³²TOKENIZER · ττ : 𝕏 → 𝕋⁵¹²FOUNDATION MODEL · ƒtransformer · 7-dim behavioral profileEmbeddingAttention × 12FFN × 12Layer NormProfile HeadFATIGUEATTNSKILLSTRESSANOMCOLLABLOADDECISION · π31actionsfrom 308 cataloguedπ : ƒ(τ(𝕏)) → 𝕐OUTPUTS · 𝕐specific actionsREST15-min breakROTATEA → B stationRETRAIN4 workers¥ IMPACT+¥12,400 / shiftVALIDATION AUC 0.737 across 21 public HAR datasets · n = 1.2M sequences · trained on 33 datasets · 15M+ actions analyzed
III · MODELIII / VIII
Chapter IV · Live Dashboard
LIVE · 998 TESTS · 21 RULES

What a real production line looks like, right now.

  • · Live demo on simulator.tangerineintelligence.ai — the same engine, running on a simulated PCB line.
  • · 21 causal rules, 998 tests, 18 streams. The numbers behind every tile are auditable.
PCB Production Line · Live
14:00
SMT ZONE
MANUAL ASSEMBLY
TESTING
FQC & PKG
SPP
P&P-1
P&P-2
REFLOW
AOI
ICT
FT-1
FT-2
FT-3
GW-01
GW-02
GW-03
Tangerine iFactory
PCB Assembly Line · Demo
Output Rate
0%
Quality Rate
0%
Avg Fatigue
0
Active Alerts
0
100/ 100Line Health
Worker Status
Read the deep dive →
daizhe@tangerineintelligence.aiIV / VIII
Chapter V · Research / IP
33 datasets · 11 patents · 9 papers
V / The model is the moat

The model is the moat. The data is the proof.

15M actions, 1.15M participants, 12 domains. 9 papers in flight, 11 active patents.

  • · 33 proprietary datasets · 15M actions · 1.15M participants · 12 domains.
  • · 9 papers in flight (target: Nature Human Behaviour, Nature Digital Medicine).
  • · 11 active patents.
DATASETS
33
12 domains · 15M actions
  • Foursquare Check-ins · 4,921,569
  • CS:GO Professional · 2,190,434
  • MovieLens 25M · 2,011,646
  • ASSISTments 2009 · 1,011,079
  • Yoochoose · ~1,000,000
see appendix →
PATENTS · DRAFTED
11
US-only · drafted
  • Behavioral state sensing
  • Sensor fusion
  • Privacy-preserving inference
see appendix →
PAPERS · IN FLIGHT
9
target Nature Human Behaviour et al.
  • Paper A · Nature Human Behaviour
  • Paper B · Nature Digital Medicine
  • Paper C · Journal of Manufacturing Systems
  • Paper 2 · Nature Human Behaviour
  • Paper 5 · Nature Human Behaviour
see appendix →
V · RESEARCH / IPV / VIII
Chapter VI · Proof
LIVE · SIMULATOR + CONSOLE
VI / Live product

The product runs today. Two public surfaces. See for yourself.

A live behavioral simulator and a live factory console. Active pilots with PCB manufacturers in the Pearl River and Yangtze deltas, deployments in progress.

Public surfaces — live
iFactory — In Conversations

PCB Tier-2 specialty manufacturers · Yangtze + Pearl River deltas

iFactory itself: in conversations with PCB Tier-2 specialty manufacturers across Yangtze + Pearl River deltas. No signed iFactory customer yet.

Status
In conversations · no paying customer
Geography
Yangtze Delta + Pearl River Delta
Segment
PCB Tier-2 specialty private manufacturers
VI · CUSTOMERSVI / VIII
Chapter VII · About
INC. 2026 · DE C-CORP
VII / About us

Two people, one bet: the worker is the last unmeasured variable.

Tangerine Intelligence Inc. — Delaware C-Corp, 2026. Operating between Berkeley, Shenzhen, and the factories we measure.

Daizhe Zou / 邹岱哲

CEO · CO-FOUNDER

UC Berkeley. Family factory background. Building the Tangerine Behavioral Model from a 21-year-old's vantage on the factory floor.

Hongyu Xu / 徐弘宇

CTO · CO-FOUNDER

Co-founder & CTO. Leads the Tangerine Behavioral Model — sensor pipeline, training, on-device inference.

TANGERINE INTELLIGENCE INC. · DELAWARE C-CORP, 2026 · BERKELEY · SHENZHEN
VII · ABOUTVII / VIII
Chapter VIII · Contact
VIII / VIII

Talk to us.

Tangerine Intelligence Inc.
Delaware, USA · Shenzhen, China
daizhe@tangerineintelligence.ai
© 2026 Tangerine Intelligence Inc.
VIII / VIII · fin.