P-01PAIN POINT
Highly variable handwriting
Different sales reps write with different styles, pressure, angles, cross-line notes, corrections, signatures, and overwritten fields. Traditional OCR often fails at the full-row level.
AI Solution Brief· GSAI · 2026 · 04 · 0017Delivered / Production
Turn a handwritten order into
structured data
ready for accounting in 3 seconds
This is one entry from Boilingwater Technology's AI solution library. For handwritten order recognition, we benchmarked general OCR APIs, multimodal large models, and a domain-specific OCR model, then delivered a hybrid production pipeline. On messy handwriting, folded paper, cross-line corrections, and overlapping fields, field-level accuracy improved from 68.4% to 96.1%.
Field-level accuracy
96.1%
1,200 complex handwritten samples
Average processing time
1.8s
End to end · P95 ≤ 3.0s
Manual review workload
-82%
Before vs. after launch
Production runtime
11 months
6 order types / 3 customers
/ 01The Problem
The real business problem
The customer is a fast-moving consumer goods distributor serving hundreds of retail outlets. Orders arrive as handwritten paper forms and must be entered into ERP for fulfillment and reconciliation. The actual paper conditions are far messier than a clean demo sample.
P-02PAIN POINT
Ambiguous field meaning
Quantity, unit price, product name, and remarks are often mixed together. Product shorthand must be normalized against business vocabulary and SKU context.
P-03PAIN POINT
Manual entry became the bottleneck
The team had to process 1,200–1,800 orders per day within a narrow window. Two operators were working overtime and still produced avoidable entry errors.
P-04PAIN POINT
Small errors created financial risk
A wrong quantity, price, or customer name affects fulfillment, reconciliation, and month-end settlement. The impact is not inconvenience; it is real financial exposure.
/ 02Live Sample
One order through the full pipeline
The sample below shows a desensitized handwritten order from a real workflow. We place the original image, visual recognition overlay, and final structured JSON side by side so stakeholders can see what the AI actually does.
INPUTOriginal handwritten order
jpg · 1408×1056
OUTPUTRecognition output
Folded paper100%
Creases shift field positions, but all 12 rows are correctly aligned.
Cross-line correction97%
Crossed-out prices and handwritten replacements are interpreted with an audit trail.
Overlapping fields94%
Remarks overlapping the price column are reassigned through structured post-processing.
/ 03How We Think
We did not bet on one model. We combined the right tools.
For production AI, we do not choose a model first. We benchmark practical options against real samples, then design a hybrid pipeline where each model handles the part it is best suited for.
APath
General cloud OCR APIs
Tencent Cloud / Alibaba Cloud / Baidu AI Cloud
- Fast to integrate
- Excellent on printed text and regular tables
- Usage-based cost works for early testing
- Weak on casual handwriting and connected strokes
- Cannot understand customer-specific product shorthand
- Field structure must still be rebuilt by the business layer
- Printed accuracy
- 98%+
- This scenario
- 68.4%
- Structuring
- Custom build
Our call · Useful as a baseline and fallback channel
BPath
Multimodal LLM reading
GPT-4o / Qwen-VL / Tongyi Qianwen VL
- Strong contextual understanding for corrections and messy forms
- Can output structured JSON directly
- Good semantic generalization across product vocabulary
- Higher cost and latency per order
- Field-level numeric accuracy can fluctuate
- Private deployment and compliance require additional design
- This scenario
- 88.7%
- Latency
- 4.2s
- Cost / 1k pages
- ≈ ¥48
Our call · Best for difficult samples and semantic correction
CPath
Domain-specific OCR model
Boilingwater · GS-OCR-Hand v2
- Fine-tuned on real customer samples
- Two-stage detection and recognition is controllable and explainable
- Lower inference cost and latency
- Needs ongoing data-loop maintenance
- New form types still need migration samples
- Does not fully solve cross-field semantic judgment alone
- This scenario
- 93.2%
- Latency
- 0.9s
- Cost / 1k pages
- ≈ ¥6
Our call · Primary path for more than 90% of daily traffic
Final Decision
Hybrid by design
The delivered system uses a domain OCR primary path, a multimodal semantic correction path, and cloud OCR fallback. Most routine orders finish in under one second. Low-confidence fields are routed to the multimodal model, and poor-quality samples or service failures fall back automatically with manual-review flags.
This structure balances accuracy, cost, latency, and controllability. The engineering principle is simple: use software architecture to turn model uncertainty into business certainty.
/ 04How It Works
An explainable, degradable, and improvable AI pipeline
We treat AI as a pipeline, not a black box. Every step has a defined responsibility, input, output, and fallback strategy.
- STEP / 011
Capture and layout normalization
Images enter through mobile, scanners, or forms. Distortion, shadow, white balance, and layout are normalized before recognition.
DocAlignDeskewShadow Removal - STEP / 022
Field detection
A layout-aware detector identifies headers, rows, columns, and field roles before recognition.
DBNet++Layout-awareRoI Routing - STEP / 033
Character recognition
GS-OCR-Hand v2 is fine-tuned on real handwritten samples. Low-confidence fields are routed forward for semantic review.
CRNN + AttentionDomain fine-tuneConfidence routing - STEP / 044
Multimodal semantic correction
For corrections, cross-line notes, and context-heavy fields, a multimodal model reads against SKU dictionaries and historical order context.
VLMRAG · SKU dictSelf-consistency - STEP / 055
Structuring and rule validation
The output is normalized with unit conversion, price-range checks, customer matching, total checks, and audit logs.
Rule engineEntity resolveAudit log
System architecture · Layered view
v2.1 · 2026.04L1 · Access
Capture and access layer
Five business channels with idempotency, rate limits, and desensitization.
Channel
Mobile app
Channel
Scanner webhook
Channel
Feishu form sync
Gateway
API Gateway · JWT
L2 · Inference
AI inference layer
Three inference paths plus a confidence-aware router.
Primary
GS-OCR-Hand v2
Semantic
Qwen-VL · LLM
Fallback
Tencent OCR
Router
Confidence router
L3 · Business
Business orchestration layer
Connects OCR output to ERP, reconciliation, and review workflows.
Engine
Field normalization
Match
Customer / SKU matching
Audit
Total validation
Workflow
Manual review queue
L4 · Data Loop
Data-loop layer
Online corrections flow back into datasets so the model improves monthly.
Storage
Sample store
Label
Online labeling
Train
Monthly fine-tune
Monitor
Drift alerts
/ 05How We Deliver
Ten weeks to launch, then a continuous data loop
We deliver AI systems as accepted, measurable engineering projects. Before launch, the work is milestone-based; after launch, the data loop keeps improving the model.
- Week 0Milestone · 01
Scenario diagnosis
Walk through the real order flow with business and IT stakeholders.
Process mapData and compliance checklistSuccess metrics - Week 1–2Milestone · 02
Data cold start
Collect 4,600 real forms and build the first training and evaluation sets.
Labeling guide1,200-sample evaluation setBaseline metrics - Week 3–4Milestone · 03
Parallel path evaluation
Run cloud OCR, multimodal reading, and custom OCR on the same samples.
Evaluation reportCost-latency-accuracy quadrant - Week 5–7Milestone · 04
Hybrid pipeline build
Build confidence routing, semantic correction, fallback, and pressure tests.
Model weightsInference service v1.0Load-test report - Week 8–9Milestone · 05
Gray release
Run AI and human entry in parallel at one warehouse for reconciliation.
Daily reconciliationReview interface v1 - Week 10Milestone · 06
Full launch
Roll out to six warehouses and sign off against KPI targets.
RunbookRollback plan - OngoingMilestone · 07
Data loop and monthly iteration
Online errors flow back into the sample store for incremental improvement.
Monthly samplesModel review notes
/ 06The Result
Turning uncertain recognition into measurable business value
Instead of vague claims, we use same-sample before-and-after metrics and customer feedback to show whether the system solved the real problem.
Metric
Before
After
Change
Field-level accuracy
BEFORE68.4%
AFTER96.1%
+27.7pp
Average processing time
BEFORE≈ 90s manual
AFTER1.8s
-98%
Daily peak throughput
BEFORE≈ 1,800 pages
AFTER≈ 12,000 pages
×6.7
Reconciliation discrepancies
BEFORE0.7%
AFTER0.04%
-94%
End-to-end cost per page
BEFORE¥0.42 manual
AFTER¥0.06
-85%
Manual review workload
BEFORE2 full-time operators
AFTER0.4 FTE review only
-82%
“Month-end reconciliation used to be our biggest headache. Since this OCR system went live, orders from six warehouses are basically scanned, structured in seconds, and written into ERP. More importantly, the models and data stay in our private cloud.
ITIT Director · FMCG distributor in South China
Sign-Off
Sign-off milestones
- UAT passed2025.06.20
- Full launch2025.06.27
- First monthly review2025.07.31
- Annual renewal2026.05.08
/ 07Where It Fits
The same pattern applies to a family of document problems
The hybrid OCR + multimodal correction + fallback pattern can be reused for forms, tickets, handwritten logs, and business documents where structure matters.
Restaurant supply chainReusable →
Handwritten hotel and restaurant order forms
Stock replenishment forms can go directly into inventory systems.
FMCG distributionReusable →
Paper return slips from retail outlets
Sales teams can scan slips into monthly ledgers.
ConstructionReusable →
Site logs and signed delivery notes
Robust handling for outdoor stains, folds, and handwriting.
Healthcare distributionReusable →
Clinic or pharmacy prescription and usage forms
Sensitive units and dosage fields can be checked against dictionaries.
LogisticsReusable →
Handwritten delivery and return notes
Signatures and remarks can be separated from operational fields.
Financial documentsReusable →
Checks, receipts, and reconciliation forms
Supports local deployment and end-to-end audit trails.
/ 08About Us
Why Boilingwater Technology
We are a software engineering and AI implementation team. Over the past three years, we have moved more than 20 AI scenarios from promising demo to stable production operation.
Our Team
Make AI work in production
/ 01
Method
- Scenario diagnosis → data cold start → path evaluation → hybrid pipeline → gray release → full launch → data loop
- AI projects are managed as engineering projects with PM, milestones, and acceptance criteria.
- No endless PoC demos; every system is delivered with measurable outcomes.
/ 02
Stack
- Vision: DBNet++ / CRNN / TrOCR / GS-OCR-Hand
- Multimodal: Qwen-VL / GPT-4o / Gemini Vision with flexible routing
- Engineering: FastAPI / Triton / vLLM / Postgres / Redis / K8s for private deployment
/ 03
Engineering
- High-throughput inference service tuning
- Data-loop platform for labeling, evaluation, drift monitoring, and fine-tuning
- Unified observability across business metrics, model metrics, and cost metrics
Let's TalkIf you have a concrete workflow
If you have a concrete workflow
AI has not solved yet, let's evaluate the right approach.
The first strategy call is free. We will unpack the workflow, judge whether AI is worth using, identify the right technical route, and provide a practical initial plan and estimate within five business days.
Business email
[email protected]Office
10F, South Tower, Kingkey Yujing Times, Longgang District, Shenzhen