Turning a hardware factory's inspection and scheduling from veteran know-how into a repeatable AI workflow
Client Profile
Pain Points
Inspection relies on human eyes
Subtle surface defects are easy to miss when inspectors tire, risking full-batch returns.
Scheduling is all experience
Orders pile up and rush orders are coordinated verbally, so deadlines keep slipping.
No data accumulation
Production data is scattered across workstations with no basis for fine-grained management.
Live Demo
How We Built It
Lightweight vision inspection
Industrial cameras and edge boxes at key stations run defect-detection models with auto-stop alarms.
Order-driven scheduling
A scheduling model over order history, deadlines and capacity generates weekly plans and auto-reschedules.
Production dashboard
Output, yield and downtime stream to a real-time dashboard viewable on the owner's phone.
Not betting on one AI — combining the right few
For a real business scenario, we don't lock in a single technical path up front. We evaluate candidate paths on the same benchmark, then design a hybrid so each path does what it's best at.
General LLM API
Public-cloud models such as GPT-4o / Qwen / DeepSeek
- ✓Out-of-the-box, minimal integration cost, cheap to start
- ✕Limited generalization to vertical domains, unstable field-level accuracy, separate compliance design
Industry fine-tuned model
Domain fine-tuning of open-source models on industry data
- ✓Grasps industry semantics, much higher accuracy, private deployment
- ✕Needs labeled data plus ongoing training and upkeep
RAG + rule engine
Retrieval-augmented knowledge plus business rule checks
- ✓Highly targeted, explainable, best latency and cost
- ✕Requires continuous data-loop investment
Fusion, not selection
What we deliver is not a single model but a fused pipeline of primary path + semantic correction + graceful fallback: routine cases are handled by the industry model in seconds; low-confidence fields are re-understood by an LLM; on service anomalies the system degrades to a general API and flags manual review — achieving accuracy, cost, latency and controllability at once.
An explainable, degradable, evolvable AI pipeline
We treat AI engineering as a pipeline, not a black box: every step has clear responsibilities, inputs/outputs and fallbacks. The five-step pipeline plus four-layer architecture below is the technical skeleton.
Ingestion & access layer
Multi-channel access with unified auth, idempotency and masking
AI inference layer
Multi-path inference plus a confidence router switching by features
Business orchestration layer
Connecting results to real business flows with rule checks and closure
Data loop layer
Every online error flows back; models iterate monthly
Tech Stack
Delivery Process
Diagnosis
Two days on-site mapping the inspection and scheduling flow to pick 3 high-value changes.
Data prep
Collected and labeled 5,000 defect samples and built the first eval set.
Build
Trained the inspection model and built the scheduling algorithm and dashboard.
Pilot
Two weeks of dual-run on one line, reconciling AI and human results.
Rollout & iteration
Expanded to all lines with monthly incremental model updates.
Before vs After
“We used to fear rush orders because scheduling fell apart. Now the system reschedules in seconds and tells us when we can deliver.”
Where It Fits
Related Cases
AI Surface Defect Detection System
Deployed AI surface defect detection to replace manual inspection, sharply cutting missed defects and rework.
ZhongshanAI Scheduling & Inventory Forecasting
Built scheduling and inventory forecasting models for a lighting factory to ease seasonal order volatility.
HuizhouPredictive Maintenance Platform
Built a predictive maintenance platform to shift from reactive repair to early warning, cutting downtime losses.