Computer vision quality inspection for a Japanese precision manufacturer

Client profile

The client is a Japan-headquartered precision electronics manufacturer producing components for the automotive, semiconductor, and medical device industries. Annual revenue is approximately ¥380B (USD 2.5B). The client operates 14 production facilities across Japan, Vietnam, Thailand, and Mexico, employing over 18,000 people. Quality inspection is a critical business function: a single defect passing through to a semiconductor customer can cost the client tens of millions of yen in warranty claims and reputational damage.

The challenge

The client's Osaka flagship facility performed quality inspection through a combination of manual visual inspection (92.4% accuracy, low throughput) and a legacy machine vision system installed in 2019 (94.8% accuracy, higher throughput but with a persistent false-positive rate that required human review). Combined throughput was 1,450 units per shift against a production capacity of 2,600 units per shift, creating a permanent inspection bottleneck.

The 2019 vision system had been trained on approximately 240,000 labeled defect images, but the labeling had been done by rotating factory workers with inconsistent methodology. The client's ML team estimated a further 400,000 professionally-annotated defect images would improve accuracy meaningfully, but internal annotation capacity was limited and the ML team lacked bandwidth for a 6-9 month annotation program.

The client had explored Japanese and Chinese annotation vendors. Japanese vendors quoted at ¥820-1,100 per annotated image, making the 400,000-image target commercially untenable at ¥3-400M in annotation cost alone. Chinese vendors were more affordable but the client had concerns about IP protection given the sensitivity of their defect signature data, and about compliance with Japan's strict export controls on certain semiconductor-adjacent technology.

Why Corpshore Vietnam

Corpshore Vietnam won the engagement based on three factors: technical demonstration of high-quality labeling on a 500-image trial dataset (99.2% inter-annotator agreement versus the client's target of 97%), our Japanese-speaking QA lead who could interface directly with the client's Japan-based ML team without English translation friction, and pricing that put the total program cost at approximately ¥140M versus ¥400M+ for Japanese vendors while satisfying the client's IP and compliance requirements.

The client's ML director specifically noted our proposal was the only one that included a hybrid annotation-plus-model-training component, offering to embed 4 ML engineers to work directly with the client's existing training pipeline rather than treating annotation as a standalone service.

The engagement

34 vision annotators in Hanoi, 6 QA specialists (including 2 Japanese-fluent senior reviewers), and 4 embedded ML engineers who work on the client's model training pipeline. Team scaled from 8 annotators at month one to 44 total specialists by month six. All annotators completed a 4-week paid training program including familiarization with semiconductor and precision component defect taxonomy in collaboration with the client's ML team. QA specialists visit the Osaka facility twice annually for calibration.

Approach and methodology

Annotation-plus-modeling integrated workflow. Rather than delivering labeled data as a monthly batch handoff, our team works within the client's ML pipeline. Annotations flow into training runs weekly, model performance improvements are measured, and defect categories where model performance lags get prioritized re-annotation attention.

Active learning loop. The vision model's own uncertainty is used to prioritize which unlabeled images should be annotated next. High-uncertainty examples get priority, low-uncertainty examples are sampled less densely. This approach reduced the total annotation budget requirement by an estimated 22% versus uniform random annotation.

Weekly Japanese-language calibration. Every Friday, our Japanese-fluent QA lead runs a 90-minute calibration session with the client's ML team and their Osaka quality inspection engineers. Ambiguous cases are debated and adjudicated. This cadence has been maintained for the full 16-month engagement without missing a session.

Results

The updated vision model reached 99.2% detection accuracy on the client's held-out benchmark by month 8 (versus 94.8% for the legacy system). Effective throughput on the Osaka production line rose from 1,450 units per shift to 2,820 units per shift by month 12, effectively eliminating the inspection bottleneck. Estimated annual value from throughput gains and reduced warranty claims: ¥1.8B against a total engagement cost of ¥180M over 16 months.

The engagement has extended to two additional client production facilities (Vietnam and Thailand). The Vietnamese facility is served by our HCMC team, providing a same-country cost and cultural advantage.

Enduring value

The annotation methodology and defect taxonomy documentation produced during this engagement have become the client's internal reference for future vision system training. The client is applying the same approach to two adjacent production lines with expected value delivery of ¥1.2B annually.