Tangerine
Intelligence
Comprehensive Overview
Detecting behavioral and psychological states from accelerometer temporal signals. One sensor. Seven dimensions. Zero cameras.
iFactory
One Sensor. Seven Dimensions.
Real-time behavioral state from wrist accelerometer
MES Tracks Machines. Nothing Tracks Workers.
Manufacturing Execution Systems monitor every machine parameter in real-time. But the most unpredictable variable on the production floor — the human worker — remains a black box. Fatigue, attention drift, and stress accumulate silently until they surface as defects, injuries, or turnover.
ILO estimates fatigue-related manufacturing losses exceed $136B annually worldwide.
In Guangdong province, manufacturing worker turnover runs 20-35% annually. Average employer cost per worker: ¥8,000-11,000/month. Every departure erases months of accumulated skill.
Precision Manufacturing
Machines only operate as well as the humans overseeing them.
Biometric Blind Spot
No existing system captures real-time worker behavioral state at scale.
The Last Blind Spot
Machines have MES, materials have ERP, warehouses have WMS. Workers have nothing.
Why Now
Labor Costs Hit Critical Mass
Chinese manufacturing labor costs have tripled in the past decade. Factories can no longer absorb inefficiency through cheap labor — every percentage point of productivity now matters.
PIPL Makes Cameras a Liability
China’s Personal Information Protection Law (2021) classifies facial recognition and video surveillance as sensitive personal data processing. Accelerometer-based monitoring avoids biometric consent requirements entirely.
Why Now
Deployment Readiness
Edge AI Chips Reach Factory Price Points
ARM Cortex-M4 processors like the nRF52832 now cost under $3 in volume. Running ML inference on-device at <50ms latency was impossible 5 years ago. Our entire wristband BOM is $5.82.
Policy Tailwinds
China’s Intelligent Manufacturing 2025 initiative and Made in China 2025 actively subsidize factory digitization. Worker-level intelligence is the next frontier after machine-level MES adoption.
We Found It in the Wrist.
This isn't a new sensor. The BMI270 accelerometer costs $0.80. What's new is the signal processing — 25 temporal pattern features extracted on-device, in real-time, with <50ms inference on a $2.40 ARM chip. The insight: you don't need cameras, EEG caps, or heart rate monitors. A single axis of motion, sampled at 50Hz, carries enough information to detect fatigue, attention drift, skill proficiency, and stress.
What does AUC mean?
What AUC 0.737 means: out of every 10 judgments, the system correctly distinguishes fatigued from normal over 7 times. Camera-based systems achieve 0.60–0.68. EEG achieves 0.75–0.80. We reach comparable accuracy with a $5.82 wristband vs. $200,000+ camera installations.
Cross-Domain Validation
Manufacturing
Fatigue onset detectable 18–30 min before quality defects
Healthcare
Stress detection AUC +0.084 over baseline (p=0.003)
Sports & Fitness
Activity recognition AUC 0.991 with dual-sensor fusion
Daily Living
3 universal behavioral dimensions: Tempo (39%), Exploration (22%), Session Arc (15%)
All datasets are from peer-reviewed published studies. We analyzed — we did not collect. Proprietary factory data collection begins with Phase 0 pilot.
14:15 — Zhang Wei, Plating Line B.
Fatigue score 0.73 (personal baseline 0.45). Plating rhythm slowed from 4.2 seconds per board to 5.1 seconds — 21% below his normal pace. Shop floor: 31°C, WBGT 27.5 — heat stress accelerates fatigue onset.
Verbal Reminder
Shift leader's watch buzzes: 'Plating B — Zhang Wei, pace dropped 21%.' Leader walks over for a quick visual check: 'Watch the copper thickness.'
Target QC
Downstream QC inspector receives notification: 'Prioritize Plating Line B output, 14:00–14:30.' Random sampling becomes targeted inspection. Each defect caught at this station saves ¥420 in rework. Caught at final test: ¥2,000+. Reaches customer: ¥5,000+.
Insert Break
If fatigue score hasn't recovered 15 minutes after the reminder, system recommends a break. Shift leader makes the call. After break: copper thickness deviation returns from ±8.3μm to ±3.1μm.
These three actions exist in every factory today. Shift leaders already do them — by gut feel. We turn gut feel into data.
Closed-Loop Feedback
10:45 — Tempo recovers to 39/min → positive feedback recorded. 11:00 — Defect rate drops to 2.1% → Level 2 effectiveness confirmed. Every decision-outcome pair trains the model.
Full Decision Spectrum
arrow_forward31 transferable decisions × 4 categories (Assignment, Intervention, Support, Development) × 4 deployment phases
iFactory: One Sensor, Seven Dimensions
A wrist-worn accelerometer that captures temporal movement patterns — not biometrics, not video, not self-reports. Motion rhythm degrades predictably with fatigue — the same way handwriting deteriorates when you're tired. Seven behavioral dimensions, starting with fatigue (validated), expanding to six more. An efficiency analysis tool that gives suggestions. Humans make decisions.
TangerineWatch (TI-WB-01)
Industrial motion sensor (BMI270 IMU) 6-axis · nRF52832 · BLE 5.0
MCU
nRF52832 (Cortex-M4F, 64MHz)
Sensor
BMI270 6-axis IMU (ACC+Gyro), 50Hz
Size
Ø16mm × 8mm, ~12g
Battery
180mAh LiPo, ≥10h at 50Hz sampling
Three Layers. One Decision.
Sensing Layer
300:1 compression
Modeling Layer
Personal baselines
Decision Layer
Prescriptive actions
Comparison vs. Alternatives
| Approach | Cost/Worker | Privacy Risk | Real-time | Accuracy | Deployment |
|---|---|---|---|---|---|
| Camera Systems | $50K-200K/line | High (PIPL — China's Personal Information Protection Law) | Yes | AUC 0.60-0.68 | Months |
| EEG Headband | $200-500 | Medium | No | AUC 0.75-0.80 | Weeks |
| ECG/HRV Patch | $200-2,000 | Medium | Partial | AUC 0.70-0.75 | Weeks |
| Self-report Survey | ~$0 | Low | No | Unreliable | Days |
| iFactory | $5.82 | Low | Yes | AUC 0.737 | Days |
† iFactory AUC 0.737 from SDK benchmark (cross-dataset validation). Range: AUC 0.52–0.94 across 21 public datasets.
Three Layers. One Decision.
Why three layers, not one end-to-end model? Because each layer solves a different problem and runs in a different place. The watch can’t store 45 days of history. The cloud can’t see raw motion for privacy reasons. The decision engine needs context that neither has alone.
Sensing Layer
感知层
Think of it like this: a step counter tells you someone walked 8,000 steps. Our system tells you their steps got 12% more irregular over the last hour, their pause-to-action transitions slowed down, and their movement rhythm shifted from the pattern they had at 8am. That’s the difference between intensity (how much) and temporal patterns (how). We extract 25 of these pattern features from the BMI270 IMU at 50Hz, entirely on the watch itself. Nothing leaves the device except a compressed 6-number summary.
300:1
Compression
<50ms
On-device Inference
Modeling Layer
建模层
The watch knows how Zhang Wei is moving right now. But is that unusual for him? The Modeling Layer answers that by maintaining a personal baseline for each worker on each process. It knows that Zhang Wei’s plating rhythm naturally slows 8% after lunch (normal for him), but a 20% slowdown means real fatigue. This layer runs on the factory edge server because it needs 45+ days of history that won’t fit on the watch.
Decision Layer
决策层
This is where we’re different from every dashboard on the market. The Decision Layer doesn’t show a score and leave the manager to figure it out. It searches through 45 possible actions (reassign, rotate, break, retrain, etc.), calculates which one saves the most money for this specific worker on this specific process at this specific time, and sends one clear recommendation to the manager’s watch: “Move Zhang Wei from Plating to Inspection after lunch. His plating copper thickness deviation increases 31% after hour 4, but his inspection accuracy stays at 98.2%. Estimated save: ¥2,100/shift (avg defect cost ¥420 × 5 defects avoided, modeled from quality-loss research).”
Sample Recommendation
“Move Zhang Wei from Plating to Inspection after lunch. Save ¥2,100/shift.”
Not all 31 decisions work on day one — some need weeks of personal data, others need months of intervention history. Here's what unlocks when, and why.
Deploy
Day 1 · Rule-driven
No history yet. Simple thresholds fire immediately — fatigue too high → swap station. Already better than guessing.
Learn
Month 2 · Prediction-driven
After 45 days we know each worker's personal baseline. Now we can predict who will decline 18 minutes before it happens.
Understand
Month 6 · Causal-driven
Enough before/after intervention data accumulates to prove what actually works — not just correlate, but cause improvement. Enables SOP revision and scheduling feedback.
Optimize
Year 2 · Optimization-driven
Cross-worker, cross-station data enables globally optimal assignment. Every person on their best-fit station at every hour. Cross-factory causal network unlocks industry benchmarks. Factory world model simulates candidate decisions before real-world implementation. Every intervention-outcome pair improves the model.
Two Layers of Value
Information alone is a cost center. Information paired with action is a profit center.
Real-time, individual-level, non-invasive awareness of every worker’s behavioral state — then act on it.
Knowing What’s Happening
Most workforce analytics stop here. They tell you: “Worker #47 has a fatigue score of 0.73.” That’s useful. But what do you do with it? The line manager still has to figure out the response. This is where 99% of industrial IoT dashboards end — a screen full of numbers that nobody acts on.
Knowing What to Do About It
iFactory’s Decision Layer doesn’t just detect — it prescribes. Every alert comes with a specific action, the expected impact in RMB, and the confidence level. The line manager doesn’t interpret data. They approve or reject a recommendation.
from type-matched interventions
Cross-dataset validation · 4 independent datasets
Knowing What to Do About It
When you know a worker's real-time behavioral state, these are the decisions a shift leader can make — organized by how far ahead they look. These aren't hypothetical. Every action here comes from our 31-decision framework, built by studying real factory floor management.
This is what we sell. Decisions with dollar signs. Why is this hard? Because it requires 45+ days of per-worker behavioral data, knowledge of which actions work for which behavioral type, and a feedback loop that learns from outcomes. You can’t skip to Layer 2 without building Layer 1 first — that’s the moat.
The Research Behind Layer 2
Every prediction must come with a specific action and a dollar impact. Otherwise it’s just a dashboard.
Beyond the Watch.
Sensing is layered. Tier 0: existing MES data. Tier 1: body-worn sensors that escalate with trust. Tier 2: environmental context that makes every reading accurate. The key insight: sensor fusion with conditional baselines.
MES Pure Software
Zero hardware. Zero cost. Immediate value.
Work Order Time Series
Tempo fluctuation patterns across shifts and days
Defect Records
Quality risk patterns correlated with time-of-day and worker assignment
Attendance & Scheduling
Fatigue risk factors from overtime, shift rotation, and consecutive days
Changeover Records
Learning curves per worker per product line
Zero cost proof of value → factory agrees to deploy hardware.
Body-Worn Behavioral Sensors
Escalating sensor deployment as trust builds.
Wrist Watch
Motion tempo, upper-body fatigue, basic behavior classification, HRV
Smart Insoles
Standing fatigue, gait analysis, weight shift patterns
Smart Gloves
Grip force, hand tremor, tool usage patterns
Chest / Shoulder Clip
Posture monitoring, bend frequency, full-body fatigue
Sensor Combination Matrix
Which station type needs which sensors.
| Station Type | Watch | Insole | Glove | Chest | Dimensions |
|---|---|---|---|---|---|
| Plating / Surface | check_circle | check_circle | — | — | 5/7 |
| Precision Assembly | check_circle | — | check_circle | — | 6/7 |
| Soldering | check_circle | — | check_circle | — | 6/7 |
| Material Handling | check_circle | check_circle | — | check_circle | 7/7 |
| QC / Inspection | check_circle | — | — | — | 4/7 |
| Packaging | check_circle | check_circle | — | — | 5/7 |
Environmental Sensors
Context that makes every body-worn reading meaningful.
ESP32-C3 · 1 node per 50m² · ¥80–200 BOM
Temperature / Humidity
SHT40Heat stress correlation. WBGT >28°C triggers fatigue acceleration.
Noise Level
MEMS Mic85 dB+ increases cognitive load 15–20%.
Air Quality (VOC)
SGP41VOC affects cognitive function. Critical in spray/injection workshops.
PM2.5
SPS30Respiratory burden in dusty environments.
Lighting
TSL2591<300 lux → eye fatigue → precision errors.
Equipment Vibration
ADXL355Separates equipment factors from human factors.
Why Environmental Sensing Isn't Optional
Behavioral features × Environmental context → Conditional baseline → Anomaly detection based on f(behavior | environment) residuals.
Not “Zhang Wei is slow today” but “Given current temperature and noise, Zhang Wei is slower than expected.”
Separating human factors from environmental factors is what makes decisions accurate.
Sensor Fusion Flow
Body-worn signals
Watch + Insoles + Gloves
Environmental context
Temp + Noise + Light
Conditional Baseline
Same worker, different “normal” at 30°C vs 22°C.
Accurate anomaly detection
Factory-Grade. Factory-Priced.
Every component is designed for harsh factory environments — dust, sweat, cleaning fluid, 24/7 shifts. Mass-production pricing from day one.
TangerineWatch
TI-WB-01The core sensing unit. One wristband per worker.
/ worker
王建国
PCB插件
14:32
8,234
步数
28%
疲劳
72%
电量
休息提醒
连续工作2.5小时
建议休息10分钟
今日班次完成
96%
效率
12,450
步数
45%
疲劳峰值
每小时活动量
Charging Rack
TI-RACK-01Charge, sync, and update — overnight, hands-free.
Edge Gateway
Two tiers. Pick what fits.
Small Gateway
≤30 workersLarge Gateway
30–100 workersEnvironmental Sensor Node
TI-ENV-01Context layer — temperature, noise, light, air quality.
Firmware OTA Strategy
Zero-touch updates. Zero bricked devices.
Distribute
Update distributed via gateway BLE broadcast
Schedule
Executed automatically during night charging (watch on rack)
A/B Slot
Dual partition design — write to inactive partition
Verify
CRC32 + Ed25519 signature verification
Rollback
Auto-rollback on failure — always bootable
Canary
5% devices first → 72h monitor → full rollout
Graceful Degradation
Every failure mode is explicitly labeled. We never fake precision.
Auto drop — no grip features
Auto drop — no standing fatigue detail
Use historical env average (marked “degraded mode”)
HR-only analysis (marked “partial data”)
Remove from real-time, supplement with MES data
Every degradation is explicitly labeled. We never fake precision.
Certification Path
Positioned as industrial efficiency tool — no medical device certification initially.
EN 62368-1 + EN 301 489 + EN 300 328
Part 15 Subpart C
GB 4943.1
IP67 (watch) / IP54 (gateway, env nodes)
Compliant
No medical device certification initially — positioned as industrial efficiency tool, not safety or health device.
Architecture
From wrist to insight in under 10 seconds.
TangerineWatch
- BMI270 6-axis @ 50Hz
- 25 temporal features extracted on-device
- 300:1 compression
Gateway
- Edge aggregation
- Local buffering
- WiFi/Ethernet uplink
iFactory Backend
- FastAPI + TimescaleDB
- 18 MES behavioral features
- Real-time scoring engine
Dashboard
- Line manager view
- Per-worker dimension scores
- Scheduling recommendations
Manager Watch
- Actionable alerts
- Specific worker + action
- Response confirmation
Data Sovereignty by Design
- check_circleRaw accelerometer data never leaves the factory premises
- check_circleOnly anonymized, aggregated behavioral scores are transmitted
- check_circlePIPL compliant: accelerometer data treated as sensitive personal information
- check_circleNo facial recognition, no video, no audio, no biometric identifiers
BLE 5.0
<100 bytes/sec per device. Ultra-low power consumption with extended range for industrial environments.
Edge Processing
25 temporal features extracted on-device. Raw data never leaves the wristband.
Privacy Architecture
Device Layer
Laplace noise injected on-device before any data transmission
Transport Layer
Per-session key rotation
Edge Layer
Cannot isolate individual from group of 5+
Cloud Layer
Only anonymized aggregates leave the factory
Re-identification Risk Waterfall
Product Positioning
Efficiency analysis tool
Safety monitoring Worker surveillance
Seven Dimensions of Worker Performance
Sequential rollout. Fatigue is validated; remaining dimensions are in development. Each dimension validated against quality and safety outcomes before adding the next.
Phase 1 Strategic Focus
Real-time fatigue scoring from IMU acceleration patterns and MES cycle time deviations.
Latency
<50ms
Data Points
15M+
王建国
TI-2024-0847
PCB插件 · A3产线
在岗186
入职天数
A
效率评分
Lv.3
技能等级
七维画像
今日班次
From Signal to Score.
How seven behavioral dimensions are actually computed — feature extraction methods, time windows, calibration, and the infrastructure that keeps models accurate.
Feature Engineering
Three representative dimensions with full computation detail.
Accelerometer 3-axis (50 Hz)
30s sliding, 50% overlap
score = 100 − clip(|z(f₀)| × 25, 0, 100)0–100 score. 100 = perfectly stable at baseline tempo.
First 3 shifts establish personal baseline. Weekly rolling update.
Remaining Dimensions
Force / Amplitude
Signals: Accel RMS + peak decay rate + motion symmetry
Window: 10s
Output: Retention % vs shift start
Posture & Ergonomics
Signals: Wrist ROM + [chest] trunk angle + [insole] CoG ellipse area
Window: 60s
Output: Risk level: low / mid / high
Collaboration Sync
Signals: Cross-worker tempo correlation + buffer accumulation + team tempo variance
Window: 15 min
Output: Team health score + bottleneck ID
Adaptability
Signals: Cross-shift / cross-station longitudinal data
Window: Day / week
Output: Changeover recovery time + shift adaptation speed + multi-skill coverage
Cold Start Strategy
From group prior to personal model in 7 days.
Group Prior
Use station-group baseline as initial prior, segmented by age, gender, and experience level.
Bayesian Update
P(personal | data) ∝ P(data | personal) × P(group prior). Each shift validated against MES. Transfer learning warm-start from similar stations.
Personal Baseline
Prior weight decays exponentially (α = 0.85^day). By Day 7, model is primarily personal data.
Cross-Station
Retain fatigue/adaptability models, only recalibrate tempo/skill. Calibration shortened to 1–2 days.
Data Labeling Strategy
Six label tiers — from free natural labels to targeted active learning.
| Label Type | Source | Method | Reliability |
|---|---|---|---|
| Output Efficiency | MES system | Direct read (hourly / per-shift output) | High |
| Quality Defects | MES / QMS | Defect records linked to worker + time | High |
| Behavior Segmentation | MES events | Changeover / downtime / shift-change as anchor points | Medium |
| Fatigue Level | Time regularity | End-of-shift 2h = positive, start 1h = negative | Medium |
| Action Categories | Initial deploy | Process engineer labels 20–50 samples, few-shot learning | Medium |
| High-Uncertainty Samples | Model output | Push to shift leader for confirmation | High |
Source: MES system
Direct read (hourly / per-shift output)
Source: MES / QMS
Defect records linked to worker + time
Source: MES events
Changeover / downtime / shift-change as anchor points
Source: Time regularity
End-of-shift 2h = positive, start 1h = negative
Source: Initial deploy
Process engineer labels 20–50 samples, few-shot learning
Source: Model output
Push to shift leader for confirmation
Each new station needs ~2 hours of process engineer annotation. After that, models self-maintain.
Model Lifecycle
Continuous improvement pipeline — develop, deploy, monitor, retrain.
Develop
- •Experiment tracking (MLflow)
- •Versioned datasets (DVC)
- •Hyperparameter search (Optuna)
- •A/B testing (shadow mode)
Deploy
- •Model registry with version + metadata
- •Canary release: 10% → 50% → 100%
- •Validate at each stage
Monitor
- •PSI drift detection (daily)
- •PSI > 0.1 → warning
- •PSI > 0.2 → auto-trigger retrain
- •Input feature distribution monitoring
Retrain
- •PSI > 0.2 (distribution drift)
- •New workers > 20% of team
- •Product/process change (MES order type)
- •Monthly incremental, quarterly full
Rollback
- •72h post-deploy evaluation
- •AUC drop > 5% → auto-rollback
- •Keep last 3 versions per factory per model
Feature Store
Single source of truth for features.
Real-time
Redis
Historical
ClickHouse / TiDB
Same feature definitions for real-time inference AND offline training — eliminates training-serving skew.
Two Axes of Intelligence
Not just “is this worker fatigued?” but “how is this worker’s fatigue trending across different processes over time?”
Intra-Shift
Real-time fatigue curve within a single 8-hour shift. Detect the 2pm crash before it causes a defect.
Inter-Week
Week-over-week behavioral pattern drift. Is this worker’s Monday performance degrading? Identify chronic issues.
Seasonal
Quarter-level skill growth or degradation trends. Track whether training investments are actually producing measurable improvement.
Process Axis
Worker x Process x Time
Recommendation
Zhang: Schedule on Process A mornings, switch to Process B afternoons. Li: Assign to Process B full-day. Wang: Keep on Process A, remove from Process B rotation.
AI Insight
Every prediction comes with a specific action and dollar impact.
The Cross Matrix
Traditional workforce analytics treat workers as static units and processes as independent variables. iFactory models the interaction — the same worker performs differently on different processes at different times. This three-dimensional view (Worker × Process × Time) is what enables prescriptive scheduling that no MES system can provide.
Traditional workforce analytics treat workers as static units and processes as independent variables. iFactory models the interaction — the same worker performs differently on different processes at different times. This three-dimensional view (Worker × Process × Time) is what enables prescriptive scheduling that no MES system can provide.
31 Management Decisions
We studied what shift leaders do when something goes wrong with a worker. Across electronics, automotive, textiles, and food manufacturing, we identified 31 management actions — then categorized them by type and mapped when each becomes available as behavioral data accumulates.
Who does what — optimal placement based on real-time worker state
Shift-start assignment
Assign each worker to their optimal station based on today's state
Mid-shift swap
Move a declining worker to a less demanding station
Bottleneck reinforcement
Send available workers to back up a bottleneck station
Rework assignment
Assign rework to workers whose current state best fits detail work
Floater deployment
Deploy floating workers where behavioral data shows most need
Changeover crew plan
Plan crew for line changeover based on skill and fatigue profiles
Worker-machine ratio
Adjust how many workers per machine based on real-time capability
Cross-line borrowing
Borrow workers from another line based on comparative state data
Temp worker scheduling
Schedule temp workers based on predicted capacity gaps
建议将王建国从A3产线调至B1产线
建议A2产线集体休息提前至15:00
Simulate Before You Decide.
A learned dynamics model of the factory. Run candidate decisions in a virtual factory. Pick the one that saves the most money — before committing real resources.
State → Transition → Reward → Planning
State (Sₜ)
Factory state vector: worker states (7 dims × N workers) + station states (capacity/defect rate × M stations) + environmental state (temp/humidity/noise) + production state (orders/deadlines).
50–200 dimensions depending on factory scale
Transition Model (fθ)
Learned state transition function: Sₜ₊₁ = fθ(Sₜ, Aₜ) + ε. Ensemble of MLPs — deterministic mean prediction + probabilistic uncertainty estimation.
Trained on historical state changes before and after each decision
Reward Model (R)
Multi-objective reward: R = w₁·output + w₂·quality + w₃·(−fatigue) + w₄·(−safety risk) + w₅·(−intervention cost).
Weights set by factory management — they decide priorities
Planning
Search for optimal decision sequence over planning horizon H (typically 4–8 hours = remaining shift). Method: CEM or MPPI with safety hard constraints.
A* = argmax Σ γʰ · E[R(Sₜ₊ₕ, Aₜ₊ₕ)]
What If We Had Let Zhang Wei Rest 30 Minutes Ago?
Roll back to state Sₜ₋₃₀
Counterfactual decision A′ = "Insert break"
Forward-simulate 30 steps via Transition Model
Actual trajectory vs. counterfactual trajectory
¥180
¥2,100
¥1,920
This isn’t hindsight — it’s free training data. Every “what if” is a training sample that teaches the system optimal decisions without real-world mistakes.
Cross-Factory Transfer
Same Industry Transfer
e.g., Plating Factory A → Plating Factory BProcess-specific transition parameters
Environmental params (different temps), personnel params (different team)
1–2 weeks of local data fine-tuning
New factory starts with prior knowledge instead of zero
Cross-Industry Transfer
e.g., Plating → Injection MoldingUniversal parts only (fatigue model, adaptability model)
Process-specific tempo model, skill model
3–4 weeks
Still much faster than learning from scratch
Progressive Intelligence Unlock
Rule Engine
Simple thresholds, already better than guessing
Prediction
Personal baselines, predict decline 18 min early
Causal
Enough intervention data to prove what works
World Model
Simulate, optimize, transfer
- chevron_rightCross-worker, cross-station optimization
- chevron_rightEvery intervention-outcome pair improves the model
- chevron_rightFactory world model simulates candidate decisions before real-world implementation
Protected by Patent TI-2026-011: Factory World Model for Decision Optimization — learned dynamics simulation, counterfactual inference, multi-objective optimization, closed-loop calibration, cross-factory transfer.
Intellectual Property & Research
Patents Drafted (11)
A competitor must break through all 4 defensive layers to replicate — each independently patented.
All 11 patents: v7 Final drafted. Filing begins Q2 2026.
warningPatent #1 must be filed before Paper B publication to preserve novelty.
auto_graphLayer 3: Causal Knowledge Graph
Every intervention creates a (state, action, outcome) triplet. Over time, these triplets form a causal network that maps which actions actually work for which behavioral states.
Research Papers
9 Internal Research PapersBehavioral Signals as the Third Modality of AI Perception
The Second Layer: Movement Patterns Predict Mental Health Where Intensity Cannot
Stress AUC +0.084 (p=0.003), Depression r=0.336 vs 0.046
Foundational finding
ARI = 0.007 between behavioral types and self-report types (what people DO has zero correlation with what they SAY)
Universal dimensions
33 datasets, ~15M actions, 3 universal dimensions (Tempo 39%, Exploration 22%, Session Arc 15%)
Competitors can copy hardware and algorithms. They cannot copy 2 years of intervention→outcome data.
Competitive Landscape
The gap between enterprise surveillance and actionable worker intelligence.
$5.82/worker
Real-time, PIPL compliant
Low Granularity
High Granularity
Key Differentiators
Single Sensor
Competitors need cameras ($$$), EEG headbands (uncomfortable), or multi-sensor arrays (complex). We use one accelerometer.
Privacy by Architecture
No facial recognition, no video, no audio. Accelerometer data processed with differential privacy (ε=3.0) and semantic compression — raw motion data cannot be reconstructed.
Edge-First
ML inference runs on-device. Raw data never leaves the watch. Only behavioral scores are transmitted.
$500B+ Lost. Zero Data.
Every factory measures machines to the millisecond. No one measures the people running them. Human behavioral factors — fatigue, errors, skill mismatch, attention lapses — drive 40–60% of quality defects and 3–5% of output loss. On $16T of global manufacturing, that’s $500B+ invisible every year. No incumbent. No standard. No data.
Value Pool (Top-Down)
Sources: UNIDO manufacturing output; ILO occupational cost estimates; factory-level validation (¥57M/1,200 workers)
Beachhead
China manufacturing (~28% global). Starting with Pearl River Delta electronics.
Ideal Customer Profile
Our Path (Bottom-Up)
Bottom-up revenue path
5 factories × 1,000 workers × $8/mo. Starting with CEO’s family factory (Sihui Fushi, 300852.SZ).
100 PRD electronics factories. Same hardware, referral-driven sales through manufacturing networks.
Automotive, semiconductor, pharma. New industry profiles, same behavioral signature. Data flywheel compounds.
Sihui Fushi Electronics
300852.SZ · ¥19.32B Revenue (2025) · 1,200 Production Workers · 3 Shifts
Based on factory data and ROI modeling — deployment pending. Analysis based on public financial data (300852.SZ annual reports) and industry benchmarks. Proposal sent; engagement pending.
Axis 1: Process × Human Factor
analyticsEach PCB production process has a different degree of human-factor dependence. The higher the percentage, the more a worker’s behavioral state (fatigue, attention, skill) directly impacts quality and cost.
Decision Framework: Built From the Ground Up
31 universal decisions — We studied shift leaders across manufacturing — electronics, automotive, textiles, food. These 31 management actions are what they actually do when something goes wrong with a worker.
20 for PCB manufacturing — Not every action applies to every industry. For PCB, 20 are high-impact — soldering quality, inspection accuracy, and plating consistency drive the most value.
7 on your Final QC line, week one — Your pilot line doesn't need all 20. We deploy the 7 that match your line's failure modes and process constraints.
4 day one, zero history needed — Reassign, Remind, Break, Handoff. These work with threshold data alone — no behavioral history required.
Axis 2: P&L Line × Time Horizon
Each P&L line item captures value at different speeds. Read the columns left to right as a timeline of what you get and when:
Instant
Week 1-2. Fatigue alerts prevent defects immediately. No model training needed.
Short-term
Month 1-3. Per-worker models stabilize. Overtime optimization begins.
Mid-term
Month 3-6. Behavioral types emerge. Team composition, hidden capacity unlocked.
Long-term
Month 6-12. Full feedback loop. Customer value, cross-process transfer learning.
| P&L Line | Ceiling | Instant | Short-term | Mid-term | Long-term | Total | % Captured |
|---|---|---|---|---|---|---|---|
| Quality cost | 1,877 | 375 | 188 | 563 | 563 | 1,689 | 90% |
| Labor efficiency | 683 | 34 | 102 | 273 | 164 | 573 | 84% |
| Overtime | 878 | 26 | 220 | 263 | 184 | 693 | 79% |
| Turnover | 100 | 0 | 0 | 60 | 30 | 90 | 90% |
| Training | 106 | 0 | 0 | 74 | 21 | 95 | 90% |
| Hidden capacity | 510 | 0 | 0 | 255 | 204 | 459 | 90% |
| Customer value | 424 | 0 | 21 | 127 | 220 | 368 | 87% |
| Operational wear | 700 | 70 | 105 | 245 | 175 | 595 | 85% |
| Safety | 200 | 40 | 30 | 70 | 40 | 180 | 90% |
| WIP capital | 250 | 0 | 13 | 113 | 100 | 226 | 90% |
| Total | 5,728 | 545 | 679 | 2,043 | 1,701 | 4,968 | 87% |
All numbers in 万元 (10K RMB).
The Intersection: Where the Money Is
When you overlay Process (Axis 1) on top of P&L × Time (Axis 2), a clear deployment strategy emerges. Not every process matters equally at every time horizon.
- Instant WinPlating + Appearance
Deploy fatigue alerts on these high-human-factor lines first. A tired plating worker deviates copper thickness by 31% after hour 4. Catching this on day 1 saves defects immediately.
¥3.75M/year from week 1
- Mid-term UnlockLamination + Solder Mask
After 45+ days of per-worker data, behavioral types emerge. Now the system can recommend: “Worker Li performs 18% better on lamination in morning shifts.” Shift scheduling optimization kicks in.
¥16.89M/year — one-third of total
- Long-term CompoundAll Processes + Cross-transfer
By month 6, behavioral models from plating transfer to solder mask (related motion patterns). Hidden capacity recovery, customer complaint reduction, and turnover prediction all compound.
¥17.01M/year — the moat deepens
Deployment Plan: Sihui Fushi Pilot
Pilot lines: Final QC + Manual Soldering
Day 1
4 actions
Rule-driven alerts on two highest human-factor lines
Month 2
7 actions
45-day baseline complete. Prediction-driven actions unlock.
Month 6
15 actions
Causal-driven actions. Full intervention→outcome loop.
Target: ¥2.82M/year recoverable value
data_explorationThe Intersection: Where the Money Isexpand_more
The key insight: Competitors who only sell Layer 1 (dashboards) can capture the “Instant” column. But the mid-term and long-term columns — which hold 68% of total value — require Layer 2 (behavioral types + action recommendations + feedback loops). That’s why Layer 1 alone is worth ¥3.75M, but Layer 1 + Layer 2 is worth ¥49.68M.
iFactory 班次报告
工厂: 四会富仕电子 · A栋
2026-04-14 · 白班 07:00-19:00
生成时间: 19:05 (自动生成)
今日执行的AI建议
| 建议 | 时间 | 结果 |
|---|---|---|
| 王建国A3→B1换岗 | 14:35执行 | 产出+15% ✅ |
| A2集体休息提前 | 15:00执行 | 下午疲劳↓12% ✅ |
| C1新人配对调整 | 09:20执行 | 技能传递+效果待观察 ⏳ |
关键事件时间线
明日建议
Business Model
Hybrid delivery: low-cost hardware for adoption, high-margin SaaS for value capture.
Hardware
Rapid Market Infiltration
SaaS Platform
Layer 1: Dashboards
Essential monitoring, fatigue alerts, and data orchestration per worker.
Layer 2: Decisions
Predictive analytics, behavioral types, and prescriptive scheduling.
90%+ Gross Margin Target
ROI Calculator
Scale your workforce to visualize the compounding revenue effect.
Current Annual Loss
$2.0M
iFactory Annual Cost
$96K
Net Savings
$1.9M
ROI
1983%
For Sihui Fushi (300852.SZ, 1,200 workers): Conservative annual value estimated at ¥26.4M (1.37% of revenue). Baseline: ¥49.7M (2.57% of revenue).
The Flywheel Effect
As sensor network density increases, the value of Layer 2 analytics grows exponentially. Predictive risk scoring becomes a mandatory compliance standard, locking in long-term enterprise contracts.
More sensors → better models → more value → more adoption → more sensors. The data flywheel compounds.
Go-to-Market
Not cold calls. Relationship-driven expansion through the manufacturing network we grew up in.
Family Factory Validation
- arrow_rightFather's factory (万坤 ihome Solutions, Dongguan) — 28 years manufacturing caps for Hennessy, Chanel, Dior
- arrow_rightPhase 0 pilot approved. Real production data, real workers, real outcomes.
- arrow_rightObjective: Prove fatigue dimension reduces defect rate on one production line.
First Deployment Target
- arrow_rightSihui Fushi Electronics (300852.SZ) — Listed PCB manufacturer, ¥19.3B revenue, ~1,200 production workers
- arrow_rightProposal and contract sent. Phase 0 data audit to determine MES traceability.
- arrow_rightOne public company validation = credibility for the entire Pearl River Delta.
Pearl River Delta Cluster
- arrow_rightExpand through father's 28-year manufacturing network in Dongguan/Shenzhen
- arrow_rightTarget: 10–20 factories in first 18 months post-validation
- arrow_rightAdditional leads: First Quality Circuit (Thailand, proposal sent), OnePlus (pitch prepared)
Regional Scale
- arrow_rightSoutheast Asia expansion (Thailand, Vietnam, Malaysia)
- arrow_rightSDK licensing to watch OEMs (OnePlus, Xiaomi, Amazfit, OPPO)
- arrow_rightSecond revenue stream: per-device royalty $0.10–0.50/watch
The unfair advantage: We didn’t cold-email our way into manufacturing. We grew up on the factory floor.
-- Founding Philosophy
Two Pillars. One Flywheel.
Tangerine Intelligence is built on two interconnected business pillars that compound over time. Each makes the other stronger.
iFactory
Perception intelligence for manufacturing
Wrist-worn sensors → 7 behavioral dimensions → 31 actionable decisions
AI Hardware Team
Full-service AI-powered hardware manufacturing
Give us your design. We handle sourcing, manufacturing, QC, logistics.
Powered by iFactory’s factory network and quality intelligence
autorenewThe Flywheel
iFactory deploys in factories → builds behavioral dataset + factory quality profiles
Quality data powers AI hardware matching — “which factory is best for this product?”
Hardware clients bring MORE factories onto iFactory (distribution channel)
More factories = more data = better models = better matching
Competitors can build a hardware sourcing platform. But they can’t build one with real-time quality intelligence from inside the factories. That’s the moat.
The Long Game
Prove
- chevron_rightiFactory in 5–10 factories
- chevron_rightValidate perception intelligence
- chevron_rightBuild the behavioral dataset
- chevron_right$480K → $2M ARR
Scale + Second Pillar
- chevron_rightiFactory in 100+ PRD factories
- chevron_rightLaunch AI Hardware Team
- chevron_rightTwo revenue streams compound
- chevron_right$10M+ ARR
Factory OS
- chevron_rightThe layer between humans and factory systems
- chevron_rightEvery factory decision flows through behavioral intelligence
- chevron_rightHardware team becomes major revenue stream
- chevron_rightIndustry-level causal knowledge graph
- chevron_right$50M+ ARR → IPO path
We don’t just measure factories. We understand them.
When you have real-time behavioral intelligence inside hundreds of factories, you don’t just sell software — you become the operating system of manufacturing.
Where We Stand
March 18, 2026
Company Founded
- check_circleDelaware C-Corp incorporated (File #10552429)
- check_circle22.5M shares authorized
March 2026
Core Research Complete
- check_circle9 internal research papers written
- check_circle33 published datasets analyzed across 12 domains
- check_circle15M+ actions, 1.15M participants analyzed
March 2026
Patent Portfolio Drafted
- check_circle11 USPTO provisional patents drafted (omnibus format)
- check_circleCovering: core algorithm, MES integration, dual-sensor fusion, behavioral signatures, integrated hardware, privacy architecture, cross-domain transfer, VBI decision optimization, fatigue budget management, BLE broadcast protocol
March 2026
SAFE Funding Closed
- check_circle$380K raised via two SAFEs at $5M post-money cap
- check_circle$80K from HK entity + $300K from individual investor
April 2026
Product Built
- check_circleTangerine SDK (AUC 0.737)
- check_circleiFactory backend (FastAPI + TimescaleDB)
- check_circleHardware specs finalized (wristband + charging rack)
- check_circleCompany dashboard, 2 demo sites deployed
April 2026
Customer Pipeline Active
- check_circleSihui Fushi (300852.SZ, family factory): proposal + contract sent — first deployment target
- check_circleFirst Quality Circuit: proposal sent
- check_circleOnePlus: pitch package prepared
- check_circleFather’s factory: Phase 0 pilot approved
Daizhe Zou
Founder & CEO
UC Berkeley, Molecular & Cell Biology
19 years old. Built entire technical stack with AI-assisted development.
Hongyu Xu
Co-Founder & CTO
UIUC, Computer Science + Neuroscience dual background
Bridges computational methods with human behavioral understanding
Now Hiring
Advisor pool: 3.5% equity reserved
“We build with Claude Code as our engineering interface — one person’s output, an entire team’s velocity.”
$500K — $1M
Pre-Seed Round. Pre-money valuation: $5–10M
Use of Funds
18-Month Milestones
One Sensor. Seven Dimensions. Zero Cameras.
We didn’t set out to build a surveillance system. We built a tool that helps factory managers understand their teams — through the physics of how people move, not the politics of how they look.
Tangerine Intelligence Inc. · Delaware C-Corp · © 2026