On the same line, same shift, output between best and worst differs 20-40%.

Machines are metered. Materials are metered.

Workers? No one is measuring.

Scroll to explore

The Blind Spot

Every factory you walk into has the same blind spot.

20-40%Output gap, same line

Same line, same shift, same SOP. The output spread between the best worker and the worst worker is 20-40%. Your MES sees the machine. Your ERP sees the material. The worker is not in either system.

ILO Global Report

$136B

ILO 2024 manufacturing labour survey, n=1,247 factories. 20-40% intra-line output spread is the median observation.

Recoverable on a 200-worker line

¥282万/yr recoverable

ILO 2024 surveyed 1,247 factories. Same-shift, same-line spread of 20-40% on output is the median, not the tail. On a 200-worker line, that is approximately ¥282万/yr recoverable, or ¥14K per worker per year.

precision_manufacturing

Machines are metered

MES reports every cycle on every station, to the millisecond. Lot codes, defect bins, takt time.

fingerprint

Materials are metered

ERP tracks every batch in and every batch out. Yields, scrap, customer returns.

visibility_off

The worker is not

No MES, no WMS, no andon records the worker. 20-40% of real work time goes unrecorded. Workers are variables, not constants.

warning

Output spread

20-40% same line

20-40% of real work

Unmeasured work time

¥282万/yr recoverable

200-worker line

How much does this invisible gap actually cost?

Worker Variance

Best vs worst on the same line: 20-40% gap. No system tracks why.

ILO 2024, n=1,247 factories. Same line, same shift, same task. The spread between top decile and bottom decile workers is 20-40% on output. No MES, no WMS, no andon tracks the cause.

Illustrative ProjectionModeled deployment value — not measured outcomes

200-worker line

single shift

20-40% spread

Same task

same SOP

different output

Cause

not tracked

worker variable

ROI ceiling

200-worker line

¥282万/年

Source: ILO 2024 · validated against same-shift line data

1,247

Factories surveyed

20-40%

Output spread observed

¥282万/yr

Per-line recoverable

Source: ILO 2024 manufacturing labour survey; recoverable value from per-line economics study (200 workers, 2 shifts).

What is your factory NOT measuring?

Machines are metered. Materials are metered. Energy is metered. The worker is not.

20-40% of real work time is recorded by no MES, no WMS, no andon.

Output = f( Machine, Material, Worker )

Machine✓MES, every cycle

Material✓ERP, every batch

Worker✗assumed constant = 1.020-40% of work time unrecorded

Behavior

Tangerine measures this

→

Action quality

→

Process output

→

Line output

→

Factory P&L

MES / ERP measure this

See the cause → prevent. See the effect → repair.

↓

CDSOLUTION

So what do you do about it?

What does the supervisor actually see at 2:15pm?

outputOutput

What the supervisor sees: a list of actions, each tied to ¥ impact.

Zhang Wei, Plating Line B. Fatigue score 0.73 (his baseline is 0.45). Plating rhythm slowed from 4.2 to 5.1 seconds per board — 21% below his normal pace. Shop floor: 31°C. Heat is accelerating the fatigue.

record_voice_over

I-01

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.'

paymentsZero costschedule5 seconds

fact_check

I-05

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+.

paymentsZero costschedule30 seconds

free_cancellation

I-03

Insert Break

If fatigue score hasn't recovered 15 minutes after the reminder, system recommends a break. Shift leader makes the call. Expected outcome (illustrative scenario): copper thickness deviation tightens significantly post-break.

paymentsLow costschedule15 minutes

app.tangerineintelligence.ai/dashboard

PCB Production Line · DemoDay Shift 07:00–19:00

2026-04-14 Tue 14:23

Throughput94.2%

Attendance96%

Avg Fatigue32/100

Live Alerts3

Factory FloorLive · 14:23:07

Today's Action Suggestions

🔄

A4 rotation: Zhang Wei → Inspection

Confidence 92%Throughput +3.1%

⏸

A5 break reminder: Team W-A03

Confidence 87%Fatigue -18pt

⚙

Line merge: A3 into A2

Confidence 78%Save 2 workers

Live Alerts

14:21Station A4-07: HR anomaly 112bpm, rest advised

14:15A5 line: 3.2h continuous work, rotation triggered

13:58GW-02 signal restored

iFactory · Sample Mockup3 gateways online22 devices connected

Latency 12ms · Uplink 2.4KB/s

↑ Live system interface (Chinese factory deployment)

Recommendation

Zhang Wei · Manual Assembly

Rotate to lighter task

¥1,92087% conf.

Validated: correctly identifies the fatigued worker 7 out of 10 times (AUC 0.737 across 21 datasets).

200-worker line economics: ¥282万/yr saved, ¥14K/worker/year, 6-12x payback. Supervisors already make these calls. We give them the math.

loop

Closed-Loop Feedback

(Illustrative scenario) After intervention: tempo recovers, defect rate drops, Level 2 effectiveness confirmed. Every decision-outcome pair trains the model.

Full Decision Spectrum

arrow_forward

31 transferable decisions × 4 categories (Assignment, Intervention, Support, Development) × 4 deployment phases

Assignment

Intervention

Support

Development

Without us

✗
Problems discovered after the damage is done
✗
Every decision is a bet based on gut feel
✗
When the veteran leaves, the knowledge leaves
✗
Every worker managed the same way

With us

✓
Warned 18–30 minutes early. The problem never happens.
✓
Simulated 10,000 times first. Then acted in reality.
✓
Knowledge becomes system. People leave, experience stays.
✓
Every worker managed according to their own rhythm.

Solution

One wristband + one gateway + one charging dock. The worker becomes measurable.

Watch BOM $6.30 at 10K units. Setup under one day.

watch

Wrist IMU

50Hz motion data · commodity hardware

→

↓

the IP · proprietary

Tangerine Intelligence Model

→

↓

Behavioral State

✓Fatigue
·Attention
·Stress
·Physical load
·Skill
·Social
·Anomaly

What it is

A model that turns commodity wrist accelerometer data into worker behavioral state. Personal baseline per worker per process. Runs on factory edge server.

What it isn't

Not a watch. Not a camera. Not a survey. The watch is just where data comes in — BMI270 + nRF52832, anyone can buy it.

Why it's defensible

Median AUC 0.737 across 21 public datasets. 45 days of personal baseline per worker = an input no competitor has.

Validation status · honest

Fatigue

Validated

AUC 0.737 across 21 public datasets

Attention · Stress · Physical load · Skill · Social · Anomaly

In development

On the roadmap, not load-bearing for the business model

Deployed via iFactory = model + TangerineWatch wristband + charging rack. Under $50/worker BOM. Plugs into existing MES/ERP. Raw data stays on-premise (PIPL compliant).

TangerineWatch (TI-WB-01)

Industrial motion sensor (BMI270 IMU) 6-axis · nRF52832 · BLE 5.0

Under $50

50HzIP67≥10h

MCU

nRF52832 (Cortex-M4F, 64MHz)

Sensor

BMI270 6-axis IMU (ACC+Gyro), 50Hz

Size

Ø16mm × 8mm, ~12g

Battery

180mAh LiPo, ≥10h at 50Hz sampling

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	Under $50/worker	Low	Yes	AUC 0.737	Days

† iFactory median AUC 0.737 from SDK benchmark across 21 public datasets (range varies by task adjacency).

System Architecture

Three layers, one recommendation

Why split into three layers instead of one end-to-end model? Each layer solves a different problem in a different place. The watch can’t store 45 days of history. The cloud can’t see raw motion (PIPL). The decision engine needs context neither of them has on its own.

sensors

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 a proprietary set of behavioral features from the BMI270 IMU at 50Hz, entirely on the watch itself. Nothing leaves the device except a compressed behavioral summary.

<100 B/s

Bandwidth

<50ms

On-device Inference

bubble_chart

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.

psychology

Decision Layer

决策层

Most workforce dashboards stop at “here’s a fatigue score, good luck.” Our Decision Layer runs the score through a 31-action playbook (reassign, rotate, break, retrain, etc., drawn from 308 catalogued factory decisions), picks the action that saves the most money for this worker on this process right now, and pushes one line to the manager’s watch: “Move Zhang Wei from Plating to Inspection after lunch — modeled save ~¥2,000/shift.”

psychologyThis is the Decision Layer

insights

Sample Recommendation

“Move Zhang Wei from Plating to Inspection after lunch. Save ¥2,100/shift.”

SpecificActionableQuantified

Progressive Intelligence Unlock

Our Decision Layer covers 31 actions — reassign, rotate, break, retrain, etc., picked from the 308 factory decisions we catalogued. But not all 31 work on day one: some need weeks of personal data, others need months of intervention history. Here’s what unlocks when, and why.

play_arrow

Deploy

Day 1 · Rule-driven

No history yet. Simple thresholds fire immediately — fatigue too high → swap station. Already better than guessing.

trending_up

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.

hub

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.

auto_awesome

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.

arrow_forward

Can you measure behavior without turning it into surveillance?

Privacy

Raw signals stay in China. Only anonymized aggregates leave.

Raw IMU and PPG never leave the watch on the worker side. Only an on-device behavioral token leaves the wrist; the gateway uplinks anonymized, aggregated features only.

Where the data sits

Raw IMU

100 Hz wrist, never leaves the watch

Raw IMU

stays on-device

+ On-device features

Behavioral token, <500 B, every 30 sec

On-device features

+ Differential privacy

Noise added before any features leave

Differential privacy

+ Personal calibration

Per-worker threshold, no raw comparison

Personal calibration

+ Aggregation

Team-level only, leaves the factory

Aggregation

PIPL Art. 29

In-China Boundary

Raw 0 days. Features 90 days. Aggregates 2 years.

Raw signal stays on-device; only anonymized aggregates cross the border.

0 days

Raw retention

90 days

Feature retention

2 years

Aggregate retention

Can one sensor really see seven things?

Dimension 08 // Product Roadmap

Seven behavioral dimensions, rolled out in order. Each has a paper behind it.

Fatigue, Attention, Skill Proficiency, Physical Load, Stress, Collaboration Sync, Anomaly. Fatigue ships first. No dimension goes live until it shows lift on quality or safety at the pilot site. 9 papers in flight, Nature Human Behaviour and Nature Digital Medicine targets.

Deployment Progress

1 / 7Dimensions Active

14%Target Reached

Strategic Focus

Real-time fatigue scoring from IMU acceleration patterns and MES cycle time deviations.

speed

Latency

<50ms

dataset

Data Points

15M+

iFactory — 员工画像

张

张伟

TI-2024-0847

PCB插件 · A3产线

在岗

186

入职天数

效率评分

Lv.3

技能等级

七维画像

今日班次

07:00上班打卡

07:15开始作业

09:30效率下降

10:00休息

10:15恢复作业

12:00午休

13:00恢复

14:32当前

What is actually inside the hardware?

layersHardware

$6.30 BOM. 5-6 day battery. 100 Hz IMU. Edge inference on a $6 chip.

Watch + gateway + charging dock. Industrial spec. Tap any layer.

Watch (TG-WATCH-SPEC-003 Rev B)

Worker side. Every worker wears one. 100 Hz six-axis IMU, 50 Hz PPG, on-device inference, BLE 5.4 uplink. The behavioral token is the only thing that leaves the wrist.

MCUnRF54L15

Cortex-M33, 128 MHz, Bluetooth 5.4

PPGMAX30101

50 Hz, red + IR + green LED

IMUBMI270

100 Hz, six-axis

arrow_right_alt

$6.30 BOM at 10K units. 5-6 day battery on a single charge. Behavioral token <500 B at 1 per 30 s.

↓

EFENGINE

What's behind a single recommendation?

A million data points. How do you turn that into one line a supervisor can act on?

12,847,293samples processed today

filter_altDecision Funnel

A million data points become one supervisor action. Here is the funnel.

Accept / reject trains the next pass

1.15M participants across 33 datasets

15M actions, behavioral patterns extracted

7 behavioral dimensions, per worker, per shift

Candidate actions filtered by causal evidence

Top recommendations with ROI tags

One line on the supervisor's screen

Every recommendation carries four tags: Action, Effect, ROI, Confidence. The model owes you an explanation.

No black box. Each line on the supervisor's screen ships with all four.

Most analytics tools stop at a score. Ours ships a recommendation: do this action, expect this effect on yield, save approximately this many yuan, with this confidence. If any of the four is missing, the recommendation does not surface.

AERC

AAction

what the supervisor should do

EEffect

expected outcome on yield

RROI

yuan saved if you take it

CConfidence

how sure the model is

If confidence falls below threshold, the line does not appear at all.

verified

200-worker line economics: ¥282万/yr saved, ¥14K/worker/year, 6-12x payback.

How many decisions does a shift leader make in a day?

45 supervisor actions, mapped across 6 industries. Behavior data unlocks 82% of them.

We catalogued every action a shift leader actually takes on the floor: 45 universal actions, validated across electronics, automotive, textiles, food, pharma, and PCB. 82% need behavior data to be made well. That is the slice we own.

Wrist IMU$6.30 BOM

11Same line, same shiftNOW

23+ Cross-shift handoff

37Behavior-data unlocked82%

45Full action set

The 82% slice is what no MES, no ERP, no camera-only stack can reach. Behavior data is the missing input.

Same hardware. Different action set per industry. Behavior signal carries across.

Decision Framework

Three filters every recommendation passes: causal, feasible, profitable.

Causal: the behavior signal actually drives the outcome, not just correlates. Feasible: the supervisor can execute it within the current shift. Profitable: the modeled yuan impact clears the cost of executing. Any candidate that fails any one filter never reaches the watch.

31Total Actions

9Assignment

10Intervention

4Support

8Development

Tier 1

Tier 2

Tier 3

Tier 4

Who does what — optimal placement based on real-time worker state

Tier 1schedulePer shift

A-01

Shift-start assignment

Assign each worker to their optimal station based on today's state

Tier 1assignment_ind

A-02

Mid-shift swap

Move a declining worker to a less demanding station

Real-timearrow_forward

Tier 2A-03

Bottleneck reinforcement

Send available workers to back up a bottleneck station

Real-timearrow_forward

Tier 1A-04

Rework assignment

Assign rework to workers whose current state best fits detail work

Per shiftarrow_forward

Tier 2A-05

Floater deployment

Deploy floating workers where behavioral data shows most need

Real-timearrow_forward

Tier 3A-06

Changeover crew plan

Plan crew for line changeover based on skill and fatigue profiles

Dailyarrow_forward

Tier 2A-07

Worker-machine ratio

Adjust how many workers per machine based on real-time capability

Per shiftarrow_forward

Tier 3A-08

Cross-line borrowing

Borrow workers from another line based on comparative state data

Dailyarrow_forward

Tier 4A-09

Temp worker scheduling

Schedule temp workers based on predicted capacity gaps

Dailyarrow_forward

iFactory · 智能建议

🔴 高优先级 · 换岗建议

建议将张伟从A3产线调至B1产线

•A3产线当前温度31°C，WBGT 27.5

•张伟连续作业2.5h，疲劳指数上升至65

•B1产线当前人员不足，温度26°C

预计产出+12%

置信度

87%

IMU数据环境传感MES产出历史模式

14:32 · 基于过去30分钟数据

🟡 中优先级 · 休息调度

建议A2产线集体休息提前至15:00

•A2产线6名工人平均疲劳指数48，高于同时段基线38

预计下午产出+8%

置信度

72%

IMU数据环境传感MES产出历史模式

14:35 · 基于过去30分钟数据

Tier 1 — Baseline

Tier 2 — Pattern

Tier 3 — Team

Tier 4 — Predictive

↓

FGMARKET

How much money are we talking about?

Market

Prescriptive analytics for manufacturing: $13B, 22% CAGR. Wearables on factory floors: still empty.

Prescriptive manufacturing analytics is a $13B market growing 22% per year. 26 vendors mapped across wearables, vision, and manufacturing SaaS. The low-cost prescriptive wearable quadrant is empty. Behavior data on the worker is the missing input, and nobody is collecting it.

This market is zero today. There is no 'worker behavioral intelligence' category. We're not capturing share — we're creating the category.

water_drop

Market sizing

Sources: prescriptive analytics market study + competitive landscape study (26 vendors across 12 wearable, 6 vision, 8 manufacturing SaaS).

$13B

Prescriptive analytics, manufacturing

$13B

Prescriptive analytics, manufacturing

22% CAGR

Annual growth, last 3 years

Competitors mapped

empty quadrant

Low-cost prescriptive wearable

Beachhead

China manufacturing (~28% global). Starting with Pearl River Delta electronics.

factory

Ideal Customer Profile

Factory size

500–5,000 workers

Industry

Electronics manufacturing (PCB, assembly, precision components)

Geography

Pearl River Delta initially, then broader China + Southeast Asia

Pain point

High turnover (20–35%/yr), quality variance, overtime costs (25–40% of labor)

Decision maker

Factory GM or VP of Production

Avg worker cost

¥8,000–11,000/month (employer total)

stacked_line_chart

Our Path (Bottom-Up)

Bottom-up revenue path

Target ARR

Year 1 — Prove$480K ARR (Annual Recurring Revenue)

5 factories × 1,000 workers × $8/mo. Starting with CEO’s family factory and a 500-worker PCB factory.

Year 3 — Scale Pearl River Delta (PRD)$9.6M ARR

100 PRD electronics factories. Same hardware, referral-driven sales through manufacturing networks.

Year 5 — Expand$50M+ ARR

Automotive, semiconductor, pharma. New industry profiles, same behavioral signature. Data flywheel compounds.

Why couldn't anyone build this five years ago?

ACT I · WHY NOW

$6.30 BOM only became possible in 2024. Foundation models for IMU only landed in 2024-25.

trending_up

nRF54L15 shipped in 2024

Nordic Semiconductor nRF54L15 hit volume pricing in 2024. Watch BOM lands at $6.30 at 10K units, with on-device inference and Bluetooth 5.4. The hardware budget did not exist a year ago.

Cost increase since 2015

gavel

Foundation models for IMU landed 2024-25

RelCon at ICLR 2025, UniMTS at NeurIPS 2024, LSM at ICLR 2025. Three IMU foundation models published inside a single year. Behavioral signal extraction from wrist motion crossed the usability threshold.

PIPL 2021

NOW

$6.30 BOM only became possible in 2024. Foundation models for IMU only landed in 2024-25.

Deployment Readiness

south_east

south_west

north_east

north_west

memory

PIPL settled the camera play

PIPL Article 28 treats facial recognition as sensitive personal information with separate consent. Worker IMU stays inside Article 29 with raw signals on-device and aggregates anonymized at the gateway.

Policy Tailwinds

Behavior validates against quality

AUC 0.832 for bad-week prediction versus 0.688 intensity baseline. +0.144 lift. Same wrist signal. Tangerine model validated on 33 datasets across 12 domains, 15M actions, 1.15M participants.

2025

MIC Target

<$50

Per Worker

IMU Silicon Cost

¥35 → ¥5.60

2018-2026, -84%

Edge AI Capability

3 → 25

Features extractable, +733%

MES Adoption

15% → 70%

Chinese factories with MES

Business Model

Hardware sold at cost-plus. Software per worker per month. Outcomes share on the upside.

Hybrid delivery: low-cost hardware for adoption, high-margin SaaS for value capture.

memory

Hardware

check_circleWatch, gateway, charging dock.

check_circleCost-plus, sold once. Lands the account.

check_circleWatch BOM $6.30 at 10K.

Focus

Rapid Market Infiltration

cloud_done

Software

High Margin

Layer 1: Dashboards

$3-5/mo

Essential monitoring, fatigue alerts, and data orchestration per worker.

Layer 2: Decisions

$8-15/mo

Predictive analytics, behavioral types, and prescriptive scheduling.

check_circlePer worker, per month.

check_circleTied to behavioral dimensions in production.

check_circleRecurring on the SaaS line.

90%+ Gross Margin Target

calculate

Outcomes share

Scale your workforce to visualize the compounding revenue effect.

On the upside1,000

200 Units5,000 Units

Projected Monthly ARRtrending_up

$8K

200-worker line baseline

$2.0M

Recoverable estimate

$96K

Per worker per year

$1.9M

Payback band

1983%

Outcomes share terms are pilot-specific and negotiated separately.

200-worker line economics: ¥282万/yr modeled recoverable. ¥14K per worker per year. 6-12x payback band.

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.

Target Margin90%+

ModelSaaS

cycle

More sensors → better models → more value → more adoption → more sensors. The data flywheel compounds.

Reference case

FQC Thailand — a representative Tier-2 PCB opportunity.

FQC = Sihui Fushi 300852.SZ Shenzhen-listed subsidiary. PCB Tier-2 specialty. Thailand operation.

SegmentListed PCB OEM

Revenue Tier~¥19B

Engagement to date: we built and delivered FQC’s marketing site (a paid build). The analysis below models the iFactory opportunity for a customer of this profile — it is a projection, not realized results. No signed iFactory deployment yet.

Axis 1: Process × Human Factor

analytics

Appearance (handling)70% Hot

Lamination / layup65% Hot

Solder mask printing55% Hot

Plating50% Hot

Surface treatment35% Cool

Open / short circuit30% Cool

Drilling25% Cool

Each 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 Leads

20 Qualified

7 POCs

4 Deploy

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).

Illustrative · Pre-Pilot Projection

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 Win
Plating + Appearance
Deploy fatigue alerts on these high-human-factor lines first. Tired plating workers show measurable copper-thickness drift late in the shift. Catching this on day 1 prevents defects before they propagate downstream.
Quality cost reduction — captured from week 1
Mid-term Unlock
Lamination + Solder Mask
After 45+ days of per-worker data, behavioral types emerge. The system surfaces shift-fit patterns — e.g. morning-shift productivity differentials by individual worker — and shift scheduling optimization kicks in.
Scheduling + overtime savings — unlocked month 2-3
Long-term Compound
All 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.
Cross-process compounding — the moat deepens

Conservative (Reference)

Conservative

¥26.42M

/year

% of Revenue1.37%

% of Net Profit20.6%

Monthly¥2.20M

Baseline

¥49.68M

/year

% of Revenue2.57%

% of Net Profit38.8%

Monthly¥4.14M

Optimistic

¥68.24M

/year

% of Revenue3.53%

% of Net Profit53.3%

Monthly¥5.69M

*Based on model estimates; actual results depend on deployment scale and factory conditions

Deployment Plan: Target Factory A 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

Worst-case floor: ¥12-15M/year (still 9.4% of net profit). Annual iFactory deployment cost for 1,200 workers: ¥1.34M. Even the worst case delivers 9-11x ROI.

data_explorationThe Intersection: Where the Money Is

expand_more

The key insight: Competitors who only sell Layer 1 (dashboards) can capture the “Instant” column. The mid-term and long-term columns — which hold the majority of total value — require Layer 2 (behavioral types + action recommendations + feedback loops). Layer 1 alone is a dashboard. Layer 1 + Layer 2 is the system that earns the recurring spend. *Tier values illustrative — calibrated against family-factory benchmarks. Actual ROI will be measured per pilot deployment.

summarizeAuto-Generated Shift Report

app.tangerineintelligence.ai/reports/shift/2026-04-14

DEMO · 演示 · DEMO · 演示

Sample report · 示例报告 · Not real factory data

iFactory 班次报告

示例工厂 · PCB产线 · 演示数据

2026-04-14 · 白班 07:00-19:00

生成时间: 19:05 (自动生成)

产出2,847件↑3.2% vs 目标

良率98.7%↑0.4%

人均效率94.2%↑2.1%

平均疲劳38/100↓5 vs 上周

今日执行的AI建议

建议	时间	结果
工号W-A03 A3→B1换岗	14:35执行	产出↑（示例） ▸
A2集体休息提前	15:00执行	下午疲劳↓（示例） ▸
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

08:15A3产线效率预警

10:30环境温度超标提醒

14:32换岗建议触发

16:00产出目标达成

18:45班次收尾

明日建议

▸建议明日白班增加B区通风

▸工号W-A03建议安排轻负荷工位（连续3天高疲劳）

iFactory · 示例报告 · Sample Report本报告为产品演示样例，非真实工厂数据

Go to market

PCB Tier-2 specialty private manufacturers, Yangtze and Pearl River deltas. Six to ten lighthouse sites in Year 1.

Anchored in PCB Tier-2: specialty, private, mid-size, technical decision-maker in the room. Lighthouse model — each pilot site delivers a reference for the next ring of factories in the same delta.

ValidationINTERNAL VALIDATION

First DeploymentPIPELINE — TARGET

ExpansionTARGET LIST

Regional ScalePLANNED

You are here

home

Family Factory Validation

arrow_rightFather's factory (万坤 ihome Solutions, Dongguan) — 28 years of precision manufacturing
arrow_rightInternal validation only — not a commercial deployment. Real production environment, real workers, real outcomes used to harden the system before customer pilots.
arrow_rightObjective: Prove fatigue dimension reduces defect rate on one production line.

Current

verified

First Deployment Target

arrow_rightTarget Factory A — Listed PCB manufacturer, ¥19.3B revenue, ~1,200 production workers
arrow_rightPitch package and contract template prepared; outreach planned via existing manufacturing network.
arrow_rightOne public-company pilot would be credibility for the entire Pearl River Delta.

hub

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 targets: First Quality Circuit (Thailand, pitch package prepared), OnePlus (pitch package prepared)

public

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

format_quote

We don’t cold-email factories. The CEO’s family runs one.
-- Founding Philosophy

↓

GHMOAT

Can someone else do this?

Tech Moat

Datasets you cannot buy. Patents you cannot design around. Hardware that ships at $6.30.

Data takes a year to collect. The model takes a year to train. That's two years no competitor can skip.

33Published Datasets

15M+Actions Analyzed

1.15MParticipants Across Published Studies

12Behavioral Domains

hub

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.

50Factories = industry-level causal network

2 yrsOf intervention→outcome data

NON-LINEAR MAPSREASONING ENGINEAUTONOMOUS DISCOVERY

science

Research Papers

9 Internal Research Papers

Paper 1 · Nature Human Behaviour

Behavioral Signals as the Third Modality of AI Perception

Paper 2 · Nature Digital Medicine

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

Paper 3

Foundational finding

ARI = 0.007 between behavioral types and self-report types (what people DO has zero correlation with what they SAY)

Paper 4

Universal dimensions

33 datasets, ~15M actions, 3 universal dimensions (Tempo 39%, Exploration 22%, Session Arc 15%)

First-Party Factory Data

Every public dataset has the wrong distribution — consumer, athletic, or lab-simulated. Our data comes from real 10-hour shifts in real Dongguan factories. Nobody else has it — not because it's hard to build, but because factories are hard to get into.

Compounding Network Effects

Every new factory makes the model better. Median AUC across 21 public datasets is now 0.737. Factory N+1 starts above Factory N's ceiling.

format_quote

Competitors can copy hardware and algorithms. They cannot copy 2 years of intervention→outcome data.

Who else is trying to do this?

Competitive

26 companies mapped. The low-cost prescriptive wearable quadrant is empty.

12 wearable, 6 vision, 8 manufacturing SaaS = 26. Cost-per-seat from Soter at $49 down to commodity surveys near zero. Behavioral wearable at single-digit dollars with prescriptive output: empty quadrant.

Deployment Cost

Data Granularity

Enterprise EHSSAP EHS, Intelex$100K+ setup

Camera + AI SystemsHoneywell, Hikvision$50K–200K/lineHigh PIPL risk

Manual MethodsSurveys, shift logs

flare

iFactory

Under $50/worker

Real-time, PIPL compliant

Low Granularity

High Granularity

Key Differentiators

sensors

Single Sensor

Competitors need cameras ($$$), EEG headbands (uncomfortable), or multi-sensor arrays (complex). We use one accelerometer.

lock

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.

memory

Edge-First

ML inference runs on-device. Raw data never leaves the watch. Only behavioral scores are transmitted.

↓

HITEAM

Who's building this?

Team

Daizhe Zou

Founder & CEO

UC Berkeley · Molecular & Cell Biology
Built the full technical and product stack.
F-1 visa holder · equity only, no salary.

Hongyu Xu

Co-Founder & CTO

UIUC · Computer Science + Neuroscience
ML pipeline architecture and sensor data processing.

Hiring

Open roles:

· Hardware engineer (embedded systems)
· ML researcher (time-series, IMU)
· BD lead (PCB Tier-2 manufacturing network)

Where are we right now?

boltNumbers

11 active patents. 33 datasets. 9 papers in flight. Console and simulator live.

check_circle

March 18, 2026

Company Founded

check_circleDelaware C-Corp incorporated (File #10552429)
check_circle22.5M shares authorized

check_circle

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

check_circle

2026

Portfolio Depth

check_circle11 active patents
check_circleUS-only · drafted · nothing filed

check_circle

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

rocket_launch

You Are Here

2026-05

First customer engagement

check_circleFQC Thailand (Sihui Fushi 300852.SZ subsidiary)
check_circleDelivered marketing site — paid build

One wristband. Seven behavioral dimensions. No cameras.

We’re not in the surveillance business. We give factory managers a way to read their teams from how people move, not from how they look.

maildaizhe@tangerineintelligence.ai

location_onBerkeley, CA / Dongguan, China

↓

Pilot conversations are open.

We are signing pilot factories now. If yours is one, talk to us.

PCB Tier-2 specialty private manufacturers across the Yangtze and Pearl River deltas. Six to ten lighthouse sites in Year 1.

Email Daizhe Read the deep dive

Measure the worker the way we measure the machine.

daizhe@tangerineintelligence.ai

That's the story.

Everything below is supporting evidence.

APPENDIX

Appendix A1 · Thesis

Why the worker.

Why the worker is unmeasured

Every machine on the line has a meter. Every kilowatt-hour, every kilogram of solder paste, every cycle of a pick-and-place head — counted, logged, traceable. The worker is the only input on the floor that the MES has nothing to say about. The supervisor knows there is a 20-40% output gap between best and worst on the same line, same shift. The supervisor cannot point at a number to explain it. Roughly a fifth of real work time is not recorded by any system at all. The variable is loud, and the instrument does not exist.

Why behavior beats survey

Self-report surveys cluster workers at ARI 0.0005 — barely above zero. Behavior clusters at ARI 0.007 on the same population — twelve to thirty-nine times tighter, depending on the dimension. The reason is simple: people answer surveys the way they want to be seen. Their hands answer the question the way they actually are. A wrist IMU at 100 Hz captures motion tempo, postural drift, micro-pause patterns, and recovery cadence — none of which the worker can fake without losing the rhythm that produces the part.

Why now

Three things landed inside eighteen months. The nRF54L15 dropped the watch BOM to $6.30 at 10,000 units, low enough that a 200-worker line is no longer a hardware question. IMU foundation models — RelCon at ICLR 2025, UniMTS at NeurIPS 2024, LSM at ICLR 2025 — turned a sensor stream into a context window. PIPL Article 29 stabilized in 2024, giving a clean architecture for raw-data localization and aggregate cross-border transfer. None of these existed when this problem was first described in the literature. They exist now.

Appendix A2 · Tech

Hardware and model specs

Wristband

PART	VALUE	NOTE
MCU	Nordic nRF54L15	BLE 5.4 · Cortex-M33 · 256 KB RAM
IMU	Bosch BMI270	6-axis · ±8 g · ±2000 °/s · 100 Hz
PPG	MAX30101	Green / red / IR · 50 Hz
Battery	180 mAh LiPo	5–6 day life · 100 Hz IMU + 50 Hz PPG continuous
Radio	BLE 5.4 to gateway	Uplink every 30 s · raw stream stays on-device
BOM	$6.30 @ 10 K units	EE only · case / strap separate

Gateway

PART	VALUE	NOTE
v1 SoC	Espressif ESP32-S3	Aggregation only · 60 watches per BLE adapter
v2 SoC	Rockchip RK3588	6 TOPS NPU · edge inference for the 7 dimensions
Backhaul	5G or WiFi 6	Aggregate features out · raw signal never leaves China
Latency	Under 30 s	Sensor sample to dashboard tile

Model

PART	VALUE	NOTE
Behavior clustering	ARI 0.007	12-39× tighter than self-report (ARI 0.0005)
Bad-week prediction	AUC 0.832	vs 0.688 intensity baseline · +0.144 lift
IMU window	30 s sliding · 50% overlap	25 temporal features per window
Personal calibration	AUC 0.62 → 0.92	45 days of individual history dominates group prior

Appendix A3 · Financials

What one 200-worker line returns.

Numbers below are for a single PCB Tier-2 specialty line, 200 workers, two shifts. No company-level revenue projection appears on this page.

ROI walk-through

LINE ITEM	VALUE	NOTE
Workers on the line	200	Two shifts · ~10-hour rotations
Output gap, best vs worst	20-40%	Same line, same shift · ILO 2024 n = 1,247 factories
Real work time, unrecorded today	20-40%	Not in MES, WMS, or andon
Savings per worker per year	¥14,000	Yield + rework + overtime + retention, summed
Annual savings, full line	¥2.82 M	200 × ¥14K
Payback	6-12×	Year-one ROI on the line, after hardware and service

Where the ¥14K per worker comes from

LEVER	ANNUAL VALUE	MECHANISM
Yield uplift	¥6,000	Fatigue-driven defect rate falls by 0.3-0.6 pp
Rework hours saved	¥3,500	Bad-week prediction routes high-risk workers off precision steps
Overtime reduced	¥2,500	Capacity recovered from the 20-40% output gap
Retention	¥2,000	Lower turnover · recruiting and training costs amortized longer

Hardware spend per line

ITEM	UNIT COST	200-WORKER LINE
Wristbands	$6.30 @ 10 K BOM	200 units · sold at cost-plus
Gateway	Cost-plus	One per shop floor
Software	Per worker, per month	Cancelable · op-ex

This page is line-level unit economics. Customer-level pricing, share-of-savings terms, and pipeline numbers are negotiated 1-1 and not published here.

Appendix A4 · Legal

Where the company stands.

Entity

Tangerine Intelligence Inc. — Delaware C-Corp, file #10552429, incorporated 2026-03-18, 22.5M shares authorized. The US entity is the sole holder of all intellectual property assigned by the founders. Operating presence: Delaware (legal), Berkeley (R&D), Shenzhen (factory work).

Founder tax election

83(b) elections on file with the IRS via Form 15620 for both founders, postmarked within the 30-day window after stock issuance. Standard four-year vest with one-year cliff. The 83(b) is the single most important paperwork moment for a US startup founder; we did it on time and we kept the receipts.

IP assignment

All founder-created IP is assigned to the US entity through Confidential Information and Inventions Assignment Agreements (CIIAA) executed at incorporation. Patent applications are drafted in the US entity's name. No founder retains personal IP claims against the company.

Trademark

Trademark applications for Tangerine Intelligence are in flight in the US and PRC. The marks are not yet registered; do not infer registration from this page.

Founder immigration

CEO is on F-1 status at UC Berkeley. The F-1 prohibits salaried employment in the US; the CEO draws no salary and is compensated only in vested equity. O-1A petition planned post-pilot completion under the extraordinary-ability-in-AI category. Family support covers living costs through this period.

PIPL posture

Raw accelerometer telemetry is classified as Sensitive Personal Information under PIPL Article 28. Raw data remains in mainland China. Only anonymized aggregated behavioral features are transmitted cross-border. Architecture and table appear in Appendix A5.

What this page does not say

No offshore intermediate. No fundraising targets, ranges, or instruments. No counsel names. Investor-facing terms are negotiated 1-1; this page exists to confirm the corporate ground is solid, not to summarize a deal.

Appendix A5 · Data compliance

How PIPL is handled.

Accelerometer = sensitive PI

Wrist-worn IMU at 100 Hz captures motion cadence rich enough that Chinese courts treat it as a biometric. Under PIPL Article 28, this is Sensitive Personal Information. The legal posture is the same as fingerprint data, not the same as a step counter.

Localization

Raw signals from the watch stay on infrastructure inside mainland China. Primary: Alibaba Cloud Shanghai. Secondary: Tencent Cloud Shenzhen. Only anonymized aggregated behavioral features leave the country, and only after the retention table below has done its work.

Cross-border path

Cross-border transfer follows the DSO (Data Security Office) Option B path: standard contract clauses with the receiving entity, security assessment for the aggregate features, no raw signal in scope. Each export is logged with the consent record that authorized it.

Position

The system is positioned as an efficiency analysis tool, not a safety early-warning system. The supervisor sees recommendations; the supervisor decides. This places the product outside the automated-decision-making regime of PIPL Article 24.

Retention table

DATA TIER	RETENTION	WHERE IT LIVES
Raw IMU stream (100 Hz)	0 days	Discarded on-device after feature extraction
Per-window behavioral features	90 days	Factory gateway · mainland China only
Aggregated team-level metrics	2 years	Aggregated · anonymized · approved for cross-border

Appendix A6 · Competitive

Twenty-six companies in the frame.

We did the scan. Twelve wearables, six vision systems, eight manufacturing SaaS. None of them sit where the iFactory product sits.

Wearable · industrial and consumer (12)

WHOOP

Oura

Garmin Health

Fitbit Premium

Apple Watch

Polar Pro

Kinetic Reflex

StrongArm FUSE

Soter Spine

Modjoul

Triax Spot-r

Reactec HAVwear

Computer vision (6)

Drishti

Retrocausal

Invisible AI

Pensa Systems

Sensar.ai

Tulip Vision

Manufacturing SaaS (8)

Tulip

MachineMetrics

Sight Machine

Augury

Uptake

Falkonry

Braincube

L2L

The empty quadrant

Three axes: low-cost (under $50 per worker), prescriptive (recommends an action, not a chart), and worker-worn (not a camera, not a server-side dashboard). The intersection of all three is empty. Soter sits at $49 per seat and is ergonomic-only. Drishti is camera-based. WHOOP is consumer-priced and consumer-positioned. The combination of $6.30 BOM hardware, prescriptive output tied to ¥ impact, and on-the-wrist deployment is unoccupied.

Appendix A7 · Research

What sits behind the model.

Three counters. Nine papers in flight. Thirty-three datasets compiled. Eleven patents active. The work behind each one is enumerated below.

Papers

in flight · Nature Human Behaviour and Nature Digital Medicine targets

Datasets

proprietary · 15M actions · 1.15M participants · 12 domains

Patents

US-only · drafted · nothing filed yet

Papers (9)

ID	TITLE	TARGET	KEY FINDING
A	Behavioral Signals as the Third Modality of AI Perception	Nature Human Behaviour	5–8 latent behavioral dimensions across 20+ datasets, 20M+ actions; median compression loss 3%; cross-modal stress AUC 0.73.
B	The Second Layer: Movement Patterns Predict Mental Health Where Intensity Cannot	Nature Digital Medicine	Movement patterns predict depression r=0.336 vs 0.046 for intensity; insomnia r=0.589 vs 0.007; intensity alone at-or-below chance for every mental-health outcome (AUC 0.45–0.52).
C	Behavioral Patterns in Production Data Predict Systemic Performance Decline Where Averages Cannot	Journal of Manufacturing Systems	Bad-week prediction in PCB lines: AUC 0.832 from temporal patterns vs 0.688 from intensity (+0.144). 3 datasets, 12,764 records.
0	The Behavioral Foundation Model	Nature Human Behaviour	Synthesis of 6 behavioral-science traditions into 6 orthogonal dimensions; 120+ references; 6 testable predictions confirmed downstream.
1	Theory-Informed Behavioral Type Identification: A Benchmark Study	Nature Human Behaviour	18 experiments across 8 psychometric datasets, 1.15M participants. Cold-start cosine similarity >0.969 from 3 signals; cross-dimensional expansion lifts R² 48.1%.
2	Behavioral Types Beyond Self-Report	Nature Human Behaviour	Behavioral and self-report typologies are independent: ARI = 0.007 across 522 participants (Eisenberg battery), IAT (N=100K), CPC18 (N=926).
3	Universal Dimensions of Human Behavioral Variation	Psychological Science	PCA across 16 datasets, 580K+ participants: 5 universal dimensions (Accuracy, Speed, Volume, Consistency, Exploration); K=3 optimal clustering; 67% type stability week-over-week.
4	The Type Paradox	Nature Human Behaviour	Types add +0.0% to prediction accuracy (94.4%→94.41%), but type-matched interventions raise success +11.6% (p<0.0001). 79–99.8% of people span multiple clusters.
5	Action Dimensions, Not Person Types	Nature Human Behaviour	33 datasets, 15M actions, 12 domains. 3 universal action dimensions (Tempo 39%, Exploration 22%, Session Arc 15%) = 76.5% pooled variance; cross-domain PCA similarity 0.654.

Datasets (33)

NAME	DOMAIN	N ACTIONS	N USERS
CS:GO Professional	Gaming	2,190,434	~500
Age of Empires II	Gaming	~1,000,000	~5,000
Clash Royale	Gaming	4,622	~1,600
DOTA 2 Pro	Gaming	34,645	~120
Go Professional	Gaming	332,397	~100
Lichess Chess	Gaming	294,000	~10,000
Lichess Bullet Chess	Gaming	22,945	~2,000
Candy Crush Saga	Gaming	93,355	6,781
OULAD	Education	106,216	4,555
ASSISTments 2009	Education	1,011,079	8,519
Duolingo SLAM	Education	200,001	901
Stack Overflow	Education	~1,000,000	~50,000
eCommerce Retail	Shopping	~12,000	~4,300
Instacart	Shopping	500,000	~50,000
Yoochoose	Shopping	~1,000,000	~200,000
Eisenberg Battery	Cognitive	522 tasks	522
Lumosity Battery	Cognitive	65,494	65,494
OpenNeuro Cognitive	Cognitive	4,551	~50
Reddit Comments	Social	~200,000	~5,000
Yelp Reviews	Social	~500,000	~50,000
OKCupid Profiles	Dating	59,946	59,946
MovieLens 25M	Entertainment	2,011,646	13,003
Last.fm 1K	Entertainment	500,000	22
SkillCraft StarCraft	Entertainment	3,395	3,395
Poker IRC	Decision	34,431	1,658
Blackjack	Decision	131,062	N/A
ITC Intertemporal	Decision	~77,000	~3,000
Foursquare Check-ins	Mobility	4,921,569	114,322
Mouse Dynamics	Process	100,447	10
CMU Keystroke	Process	224,400	51
Eye Tracking	Process	~292,000	169
Crypto Trading	Finance	1,000	N/A
OSF Lifespan Battery	Cross-Task	200×4 tasks	200

Patents (11)

11 active patents across behavioral sensing, sensor fusion, and privacy-preserving inference. US-only, drafted; none filed yet.

Appendix A8 · Risk

Three things that could go wrong.

Adoption risk

Factory floors are conservative. A line manager who has run the same shift for fifteen years does not adopt a new device because it is clever. The wristband adoption story has to clear three bars: it must not slow the worker down, the supervisor must see a number that maps to ¥ on the same shift, and the factory owner must be able to cancel it next month if it does not. We address each bar directly: the watch ships at the worker's existing wrist, the dashboard tile shows ¥ impact under each recommendation, and the software is per-worker-per-month op-ex rather than capital equipment. We expect the first two factories to take longer than the next eight.

Hardware supply risk

The watch BOM relies on the nRF54L15, the BMI270, and the MAX30101 — all single-vendor parts from Nordic, Bosch, and Analog Devices respectively. None has a drop-in replacement at the price-performance point we operate. A shortage on any one part stops shipment. We mitigate with a 90-day rolling buffer at the contract manufacturer once the first pilot wraps, dual-sourced casework and battery, and a documented v1 to v2 gateway migration path that decouples watch hardware from the inference layer. The watch design freezes at the part level only when the factory pilot returns clean MOQ burn-in data.

Regulatory risk

PIPL is the live regulation. Three secondary motions could change posture: tightening of biometric SPI handling under PIPL Article 28 enforcement, a new labor-law restriction on workplace monitoring devices, and changes to the DSO Option B cross-border standard. The architecture is built to absorb the first — raw data already stays in mainland China, the cross-border path carries only aggregated features. The second is addressed by positioning the product as an efficiency analysis tool, not a safety system, and by routing every decision through a human supervisor. The third is the residual: if DSO standards shift, the aggregation pipeline gets re-certified; the underlying signal path does not change.