Operational Challenges and Solutions for Implementing AI Visual Inspection at Japanese-Affiliated Factories in Thailand

Operational Challenges and Solutions for Implementing AI Visual Inspection at Japanese-Affiliated Factories in Thailand

Introduction

When introducing AI visual inspection at factories in Thailand, issues such as false detections caused by changes in lighting conditions, seasonal and line-to-line variation, and resistance from on-site staff toward the new operation tend to surface only after full-scale operation begins. For production engineering and quality control managers considering inspection automation, this article organizes decision-making criteria based on concrete case examples, covering everything from how to select target processes for a PoC, to fixing the imaging environment, to building on-site consensus on response flows for false detections. After reading, you will be able to assess whether your own inspection process is suited to AI visual inspection in terms of imaging conditions, defect types, and sample size, and to consider what needs to be confirmed when expanding to other processes.

In Thailand, labor shortages combined with challenges in retaining local staff have heightened interest in AI visual inspection as a means of eliminating the person-dependent nature of visual inspection. As Japanese vendors expand their local presence and collaborate with local partners, the barriers to adoption are gradually lowering.

Challenges of Labor Shortages and Inconsistent Inspection Quality

At many factories, the retirement of skilled inspectors and a lack of successors have led to a shrinking inspection system, with visual inspection skills failing to be passed on to the next generation. Judging fine scratches or color unevenness relies heavily on experience, and when personnel change, the pass/fail criteria for the same product can shift. At Thai sites, local staff retention is often a challenge, and it is not uncommon for inspectors to be replaced every few months, placing a heavy burden on the production engineering department, which must retrain staff each time.

Variations in judgment criteria manifest as differences in pass/fail results when different inspectors examine the same lot, or even as inconsistent judgments by the same inspector depending on the time of day or fatigue level. Because this person-dependent nature is difficult to quantify in inspection records, it also becomes difficult to later identify the causes of missed defects or excessive scrapping. Even when visual inspection manuals are put in place, the final line of judgment often still depends on the inspector's own sense, and documentation alone cannot fully resolve this.

If left unaddressed, variation in judgment criteria leads to increased costs from handling claims from shipping destinations and internal re-inspections, creating a vicious cycle in which the quality control department spends excessive time investigating causes. Training new inspectors takes a certain amount of time, during which the accuracy of pass/fail judgments remains unstable. It should also not be overlooked that when inspectors are added during periods of increased production volume, the mix of staff still in training can further lower overall judgment accuracy.

This difficulty in skill transfer and the resulting variation in judgment are at the core of the motivation to automate the inspection process and quantify its standards. The expectations placed on AI visual inspection hinge on whether this person-dependent nature can be replaced with a systematic mechanism.

Expected Benefits Before Implementation

Shortening inspection time and standardizing judgment are the central effects expected before introduction. The underlying idea is to replace pass/fail criteria that previously varied by inspector with quantified thresholds, aiming for a state in which the same judgment is reached regardless of who is in charge. Because inspection records are automatically retained, a shortening of the time required to train new staff is also expected. Previously, it took time for new staff to acquire a sense of judgment by working alongside senior inspectors, but by using the images judged by AI and their underlying rationale as training materials, it becomes possible to concretely show what is being observed to determine a defect.

Reducing dependence on skilled inspectors is another reason production engineering departments consider adoption. Even when inspectors retire or are transferred, the judgment criteria themselves remain within the system, which is expected to prevent situations in which person-dependent skills are lost without being passed on. There is also the expectation that, as inspection process records accumulate as data, trends in defect occurrence can be analyzed afterward and used to improve the manufacturing process itself.

When estimating the effects, it is essential to decide beforehand which of inspection time, miss rate, or training period should be prioritized for measurement. If introduction proceeds without determining priorities, it becomes difficult to later judge whether the PoC results were good. All of these effects presuppose that the imaging environment and defect definitions are properly established, and when sharing expected values with on-site staff, it is also necessary to explain which conditions, if they fluctuate, make the effects harder to achieve. Settling on measurement items during the PoC (Proof of Concept) stage makes it easier to verify effectiveness after full-scale rollout.

How Was the Implementation Carried Out?

How Was the Implementation Carried Out?

Rather than rolling out to all lines at once, the common approach is to conduct a PoC targeting a single process where imaging conditions are more likely to remain stable, agree on the on-site pass/fail criteria as numerical targets, and then expand gradually. Discrepancies in expectations depend largely on the precision of this initial consensus-building process.

Selecting Targets for PoC Validation

Choosing a process for the PoC where the imaging angle and workpiece material are easy to fix, and where the types of defects can be narrowed down to a small number, makes it less likely that variations in lighting conditions will contaminate the verification results. Lines where multiple materials are mixed, or processes where the orientation of the workpiece changes during transport, tend to run into trouble at the stage before AI judgment accuracy can even be evaluated, because the imaging conditions themselves are not stable. Conversely, in a process where a single part can be stopped in front of the camera in the same posture for imaging, it becomes easier to learn the difference between defects and non-defects, and the PoC results come closer to the accuracy achieved after full production deployment.

As a rough guide for data volume, choosing a process where 500–1000 images of good products and about 50–100 images per defect pattern can be collected makes it easier to secure the sample size needed for verification. In processes where defects occur infrequently, it takes longer to gather this number of images, which affects the estimate of the PoC period. When selecting the inspection target, it is also necessary to check how clearly defects such as scratches, foreign matter, and color unevenness can be classified based on existing defect records from the existing records, and to decide to postpone processes where the classification is ambiguous.

For factories with multiple lines, narrowing the initial PoC target down to the one line with the best-equipped imaging environment, and then rolling out the insights gained there to other lines, is an effective order for avoiding confusion on the shop floor. At the stage of narrowing down the target process, it is also worth checking whether the imaging and judgment processing time can fit within the takt time, as this makes it easier to directly compare the PoC results against the speed requirements of actual production operation. If the selection is wrong, it becomes impossible to separate whether the issue lies with lighting, defect definitions, or process selection, and identifying the cause takes longer.

Interviews with Floor Staff and Expectation Alignment

Before the PoC, it is necessary to interview inspectors individually about what criteria they use to judge defects, using their own words, and to document this. If the process proceeds with only subjective expressions such as "the scratch is noticeable" or "the color is pale," there is no clear target against which to align the AI's judgment criteria, and dissatisfaction such as "this isn't the judgment we expected" tends to arise after the PoC. It is effective to translate these expressions into measurable terms as much as possible—such as scratch length and depth, or the extent of color unevenness—and to work with the inspectors to select samples close to the boundary line. Conducting these interviews not only with veteran inspectors but also with newer inspectors whose judgment criteria are not yet fixed, and comparing where the two groups' perspectives diverge, broadens the range of boundary samples that can be used for setting the AI's thresholds.

In parallel with these interviews, it is also necessary to agree with the shop floor on numerical targets for what the PoC will measure. Both the production engineering department and the inspectors should confirm how much miss rate is acceptable, how much double-checking work due to false positives the floor can tolerate, and how the sample size and period used for verification should be set. If this agreement is vague, even good PoC results may fail to convince the shop floor, becoming a source of disagreement when deciding on full-scale rollout.

In managing expectations, it is also important to communicate in advance that the AI will not completely replace inspectors, but will function as a first-line filter for judgments, with humans confirming judgments near the boundary line—this helps avoid excessive expectations or distrust on the shop floor. It is also essential to compile the agreed numerical targets and judgment criteria into a single document that stakeholders can refer back to during the PoC period.

What Actually Happened on the Floor?

What Actually Happened on the Floor?

When the judgment criteria agreed upon in the PoC were applied as-is to the production environment, discrepancies specific to the Thai factory's imaging conditions and on-site operations surfaced one after another. Unexpected situations have overlapped and emerged across three aspects: changes in lighting, environmental differences by season and by line, and how inspectors perceive the results.

False Detections Due to Lighting Condition Changes

False detections caused by fluctuations in lighting are a problem that tends to surface when the judgment criteria agreed upon in the PoC are transferred as-is to the production line. While the PoC often involves collecting images in an environment with fixed imaging conditions, on the production line, ceiling lighting and external light mix together, and the appearance of even the same good product can change depending on the time of day and weather. When such environmental differences change how reflections appear on the workpiece surface, using the thresholds agreed upon in the PoC as-is can result in both cases where a good product with strong reflection is over-detected as defective, and cases where a defective area is hidden by reflection and missed. This is a phenomenon that can generally occur on sites where the imaging environment is not constant.

When isolating the cause, it is necessary to first cross-reference the time of day when misjudgments occurred with the captured images, to check whether changes in lighting conditions are affecting the judgments. In cases where the defect's own characteristics have not changed, but a gap has arisen between the appearance the AI learned and the appearance due to fluctuations in the imaging environment, revisiting the lighting environment before additional training is more likely to lead to improvement. Neglecting this isolation process can lead to mistaking a lighting problem for insufficient accuracy, resulting in unnecessary repeated additional training.

Additionally, it should be recognized as a phenomenon that can generally occur at sites with fluctuating lighting environments that changes in sunlight conditions and in-factory illuminance change the imaging conditions, causing a shift in the judgment threshold. By aggregating misjudgment logs by time of day and weather, making biases in over-detection or missed detections visible, it becomes easier to judge whether measures to block external light are necessary.

Discrepancies in Judgment Criteria Due to Seasonal and Production Line Differences

When humidity and the strength of external light inside the factory change between the rainy season and the dry season, the way the workpiece surface appears under the same lighting equipment can change on a monthly basis, potentially causing a seasonal shift in thresholds separate from the time-of-day bias observed in false detections due to changes in lighting conditions. During periods of continuous clear weather in the dry season, external light tends to enter strongly, so thresholds set based on images from that period may lead to increased over-detection; conversely, once the rainy season begins, external light weakens and the contrast of defective areas decreases, which could cause a shift toward missed detections. If the PoC verification period is limited to only a few weeks, the training data may include conditions from only the dry season or only the rainy season, and judgment accuracy could break down once the other season arrives.

Differences between lines create a similar problem. If the transport speed and degree of vibration differ between lines, the degree of blur during imaging can change even with the same camera and lighting settings, and a case can be envisioned where judgment accuracy, which was stable on the line where the PoC was conducted, drops as soon as it is rolled out to a different line. Wear on fixtures and slight shifts in installation position can also be factors that change the imaging distance and angle.

To separate the two variable factors of seasonal and line-to-line differences, it is effective to aggregate misjudgment logs by month and by line separately, checking whether the bias leans toward over-detection or missed detections. Countermeasures should be considered separately depending on the cause—for example, blocking external light if over-detection is biased by season, or inspecting transport and fixtures if missed detections are specific to a particular line. Designing the system on the premise that it can be adjusted for each condition, rather than operating with a single fixed threshold across all lines and all seasons, provides a clue for detecting discrepancies early after full production deployment.

Floor Staff Resistance to Operational Changes

When double-checking occurs—where inspectors visually re-confirm workpieces that were judged defective due to over-detection—doubts about the very significance of AI adoption tend to arise on the shop floor. There were also cases where the judgment results and statistical values displayed on the operation screen were hard to interpret, which ironically extended the time from opening the screen to completing the re-confirmation. As confirmed with the false detections caused by changes in lighting conditions, when inspectors observed the phenomenon of judgments splitting for the same good product depending on the time of day, distrust toward the AI's judgment itself tended to solidify first, becoming a factor that prevented subsequent over-detections from being accepted as a normal tuning process.

A major source of resistance was that veteran inspectors, in particular, showed strong resistance to having judgments they had made for years based on experience and intuition replaced by system thresholds, and they tended to react as if their skills were being denied. This resistance could not be resolved through a single round of operation training alone; it required the accumulated effort of reviewing misjudged images on the spot together with inspectors and explaining, each time, why that particular judgment occurred. If only the thresholds were continuously adjusted without sharing the basis for judgments, the shop floor would end up merely following the operation without understanding the reasons for the changes, and while the resistance might appear to subside on the surface, it would not be fundamentally resolved.

On lines where inspectors rotated by shift, a problem also emerged in which the reasons for judgments already explained in a previous shift were not passed on to the next shift, and as the same explanations were repeated, the shop floor's response gradually became perfunctory. Keeping handover content as written records rather than verbal communication, so that anyone checking could arrive at the same explanation, proved essential for shift-based operations.

What Changed Before and After Implementation?

What Changed Before and After Implementation?

After addressing the discrepancies in lighting conditions and the resistance on the shop floor, clear differences emerged compared to before introduction in terms of inspection accuracy, work time, and how the shop floor received the system. The changes were reflected not only in the numbers but also in how inspectors approached their judgments.

Changes in Inspection Accuracy and Work Time

To confirm the effects set out before introduction—shortened inspection time and uniform judgments—it is necessary to decide in advance what criteria to use for comparison once the adjustment of lighting conditions and the operation of the misjudgment log have settled in. Indicators such as the number of over-detections, missed detections, and double-checks cannot form a valid basis for comparison unless lighting conditions are fixed and threshold adjustments by season and by line have been completed; the effect cannot be judged from the numbers of the early introduction period alone. A posture is required in which the misjudgment log is aggregated on a monthly and per-line basis, causes are isolated, and evaluation is conducted only after countermeasures have been layered in.

Regarding work time, even if the time required for the inspection itself is shortened, the time required to handle over-detections will not necessarily drop to zero. In processes that retain an operation where humans confirm borderline judgments, the inspector's work shifts from "visually checking everything" to "checking only what the AI has flagged," fundamentally changing how time is used. In terms of uniformity of judgment, as the system approaches a state where pass/fail decisions are made by the same criteria regardless of which inspector is on duty, this also leads to shortening the period needed to align standards during new-employee training.

When evaluating the effect, quantitatively setting at the PoC stage which of inspection time, miss rate, or training period will serve as the comparison basis becomes the foundation for interpreting post-introduction figures. To align the comparison targets, it is necessary to retain pre-introduction records such as the number of items processed per inspector and the re-inspection rate as numerical data; in processes where no such records remain, the post-introduction change can only be explained impressionistically. Additionally, since the period required for the effect to stabilize varies depending on line conditions, it is practically advisable to avoid directly applying the improvement margin observed in one process as the target value for another process, and instead to set a measurement period for each process individually to confirm results.

Changes in Floor Staff Perception

By switching to an operation that uses the judgment log as educational material, it became possible to compile misjudgment images and the reasons for those judgments into a list for use as training materials for new inspectors, increasing opportunities to explain judgment criteria without relying on the intuition of veteran inspectors. By continuing the accumulated practice of reviewing misjudged images on the spot with inspectors and explaining the basis for each judgment, it is expected that inspectors will shift from a state of "not knowing why it stopped" to a mindset of using the system while anticipating conditions, thinking "over-detection is likely to occur under these conditions."

Even veteran inspectors who had shown strong resistance became more receptive to threshold adjustments—not as a denial of their own skill, but as a process of translating skill into a system—once they witnessed how the borderline samples they themselves had pointed out were reflected in additional learning and actually improved judgment accuracy. Establishing an operation in which handover content at shift changes is kept as a record also helps reduce situations where inspectors after a shift change repeatedly demand the same explanations.

When checking for changes in how the system is received, rather than surveying inspectors, it is more realistic to observe changes in the frequency and content of misjudgment reports. In the early stage, reports of "not knowing why it stopped" are common; in processes where, as operation progresses, these reports shift to include predictions such as "over-detection is likely to occur again during this time period," it can be judged that the shop floor has become able to explain the system's behavior in its own words. In processes where this change is not observed, the frequency of sharing the judgment log and the level of detail in explanations need to be reviewed.

What Lessons Were Learned From This Case?

What Lessons Were Learned From This Case?

Neither the discrepancies in lighting conditions nor the resistance on the shop floor could be fully resolved through after-the-fact responses, and whether or not the premises were firmly established at the PoC stage made a substantial difference to the burden after full-scale operation began. Two points—fixing the imaging environment and reaching agreement on the response procedure for when false detections occur—serve as the branching point determining whether expansion to other processes is feasible.

The Importance of Fixing Lighting and Environmental Conditions in Advance

Adopting dedicated lighting, blocking out external light, and introducing fixtures that lock the camera position form the foundation for preventing both misdetections caused by changing lighting conditions and shifts in judgment criteria across seasons and lines. If a production line is left in a state where ceiling lighting and external light mix together, and additional training is repeated without addressing this, the training data itself will end up incorporating fluctuations in the imaging environment, leading to a recurring cycle in which adjusting the threshold each time only produces new discrepancies at a different time of day or season. Setting up an imaging booth structured to block external light and then introducing fixed lighting makes the appearance of good and defective items far less likely to fluctuate, whether within a single day or across seasons.

Fixtures also need attention, since vibration during transport or even slight deviations in placement position can change the imaging distance and angle. The structure should allow the workpiece to be stopped and photographed in the same posture and position every time. This work of fixing the environment should be planned from the stage of selecting the target process for the PoC, with an eye toward eventual rollout across the entire production line, and it is desirable to keep the imaging booth used in the PoC as close as possible to the lighting conditions of the actual production line.

Keeping to the order of considering additional training as a matter of defect feature values or threshold issues—only after misjudgments occur despite a fixed environment—helps minimize the time spent isolating the cause. Even with a fixed imaging environment, light output can decline over a span of months due to the aging or dirtiness of lighting equipment, so it is also essential to periodically inspect illuminance and keep records of any deviation from baseline values. Keeping inspection records makes it possible to quickly determine, when judgment accuracy drops, whether the cause is deterioration of the lighting equipment or a change in defect trends.

Agreeing on a Response Flow for False Detections with the Floor

It is necessary to document, at the PoC stage, who checks what and through what procedure when a misjudgment occurs, and to reach agreement on this with the field. Even after the environment has been fixed, misjudgments will not drop to zero, so if a response flow for when they occur has not been established, the complaint of "in the end, we just have to check it ourselves" will resurface.

It is recommended that the response flow be built in the order of accumulating misjudgment logs, classifying causes, and then applying additional training or a rule-based filter.

  1. Record images where over-detection or oversight occurred.
  2. Classify whether the cause is a deviation in lighting, season, or fixtures, or insufficient learning of defect features.
  3. If the cause is environmental, inspect the lighting and fixtures; if the cause is insufficient learning, add boundary samples and retrain.
  4. Feed the results of the response back to the inspectors and check weekly whether misjudgments of the same classification have decreased.

It should also be decided who is responsible for each step. Agreeing in advance on a division of roles—inspectors handling the initial classification of images, the production engineering department handling cause isolation, and the vendor or in-house system staff handling the implementation of additional training—prevents delays in response when misjudgments occur. Rather than letting the agreed-upon content end as a verbal understanding, keeping it as a document that can be referenced even during shift changes helps push the operation toward reporting anomalies proactively.

What Should Be Checked Before Applying This to Other Factories or Processes?

What Should Be Checked Before Applying This to Other Factories or Processes?

Fixing the imaging environment and agreeing on a misdetection response flow cannot simply be carried over as-is to a different process with different lighting conditions or defect types. The precision of the rollout depends on work done process by process: checking the imaging conditions, defect visibility, and required sample count, and individually rebuilding agreement with the field for each.

How to Determine Whether Your Inspection Process Is Suitable for AI Visual Inspection

Whether the imaging conditions can be fixed is the first condition to check. Whether a process is suited to AI visual inspection differs greatly depending on whether fixtures can keep the workpiece's posture, angle, and distance constant, or whether the orientation of the workpiece changes during transport. Lines where multiple materials or multiple shapes are mixed together tend to have imaging conditions that fluctuate easily, and should be given lower priority as PoC targets.

It is also necessary to check whether defect types can be clearly classified. Defects such as scratches, foreign matter, and color unevenness, where existing defect records can explain the boundary lines, are easy to have the AI learn; defects for which only the inspector's subjective impressions remain on record require starting from the work of translating judgment criteria into numerical values.

Securing a sufficient amount of data is also something to assess. It should be confirmed in advance whether 500 to 1,000 images of good items and roughly 50 to 100 images per defect pattern can be collected, and for processes where defects occur infrequently, the time required to gather this number of images needs to be included in the PoC plan.

Whether the imaging and judgment processing fits within the takt time is also an essential condition. If the processing time slows down the speed of the entire inspection process, either the hardware configuration needs to be reconsidered or the target process needs to be re-selected. Checking each of these conditions one by one, and deciding to postpone any process lacking even one of them, serves as the standard for safely advancing rollout to other processes.

How to Build Consensus with Floor Staff

After selecting the target process, the practical approach is to document three points—numerical targets, measurement methods, and the agreement period—before starting the PoC, and to set up a forum where both field staff and the production engineering department can confirm them together. How far the oversight rate and over-detection rate will be tolerated, and how the number of good/defective item images and the period used for verification will be set, should be translated into concrete figures matching the conditions confirmed during the selection of the PoC target process. Once the numerical targets are set, a weekly cycle for checking misjudgment trends with the field should be incorporated.

Items to check in the weekly reviewContent to confirm
Number and type of misjudgmentsWhether over-detection or oversight is increasing
Time of occurrence / lineWhether there is a relation to lighting or transport conditions
Response statusWhether additional training or a rule filter was applied
Field reactionWhether the burden of double-checking work is increasing or decreasing

Continuing this review from the PoC period through the early stage of full operation, and accumulating records, makes it easier to reproduce the same agreement process when rolling out to a different process. From the outset, it should be assumed that the agreed-upon content will not end as a verbal understanding but will be kept as a document that can be referenced even during shift changes. If the production engineering department decides on numerical targets alone and simply notifies the field, the sense of conviction regarding borderline judgments tends to weaken, which can result in the field treating the weekly review itself as a mere formality. Having both veteran and novice inspectors present at the review and cross-checking their differing perspectives on judgment each time helps maintain the precision of the agreement.

Frequently Asked Questions

Frequently Asked Questions

We will review the rough timeline for implementation and how to proceed with improvements when false detections occur.

Q1: How Long Does It Take to Implement AI Visual Inspection?

The PoC takes 2 to 3 months, and pilot operation takes 3 to 6 months, so including the time required for full-scale deployment, a total of several months to about half a year should be expected. The duration depends on how fixed the imaging conditions are and how frequently defects occur, and for processes where it takes time to collect data on good and defective items, the PoC period will be extended. In addition, to check for seasonal shifts in judgment criteria, it is better to secure a verification period that spans both the dry and rainy seasons, as this helps avoid having to redo the work after moving to full-scale operation. When expanding to other lines, the agreement process established on the first line can be reused, which may shorten the period for the second line onward; however, please note that if the lighting conditions or line configuration differ, a similar length of verification period will be required.

Q2: How to Improve When False Detections Are Frequent?

Classify whether the issue is over-detection or missed detection, and determine whether the cause is an environmental factor or insufficient learning. If the cause is misalignment of lighting or fixtures, start by inspecting the imaging environment; if it is due to insufficient learning of defect features, start by adding samples together with a rule-based filter—this approach leads to improvement.

Summary

Summary

Whether AI visual inspection can be stably operated at a factory in Thailand depends on fixing the imaging environment and reaching agreement on the false-detection response flow. In the PoC, the starting point is to select a process where imaging conditions are easy to fix and defect types are clear, and to work with on-site staff to translate judgment criteria and numerical targets into concrete terms. After full-scale deployment, it is unavoidable that judgment criteria will shift due to fluctuations in lighting, environmental differences between the dry and rainy seasons, and differences in conveyance conditions between lines, so it is essential to prepare the imaging environment in advance using dedicated lighting and fixed fixtures. When a misjudgment occurs, respond in the order of accumulating logs, classifying the cause, and then either inspecting the environment or conducting additional learning; dividing responsibilities among inspectors, the production engineering department, and the system team makes it easier to prevent on-site complaints such as "in the end, we have to check it ourselves." When expanding to other processes, it is also effective to individually check the imaging conditions, defect visibility, and the number of samples required, and to rebuild agreement with on-site staff through weekly reviews. Please start by selecting one target process and taking stock of whether your company's inspection process meets these conditions.

References

  1. www.meti.go.jp/policy/it_policy/jinzai/AIguidebook_gaikan_ryohin_FIX.pdfAccessed November 20, 2026
  2. www.meti.go.jp/policy/it_policy/jinzai/AIguidebook_gaikan_furyo_FIX.pdfAccessed November 20, 2026
  3. www.meti.go.jp/policy/tech_evaluation/c00/C0000000R03/211208_regulation_1st/reg…Accessed November 20, 2026
  4. www.dcce.go.th/wp-content/uploads/2025/05/NAP-THAILAND-2024-English-Edition.pdfAccessed November 20, 2026
  5. www.thailand.go.th/public/issue-focus-detail/labour-minister-outlines-three-key…Accessed November 20, 2026

Author & Supervisor

Yusuke Ishihara

Yusuke Ishihara

Started programming at age 13 with MSX. After graduating from Musashi University, worked on large-scale system development including airline core systems and Japan's first Windows server hosting/VPS infrastructure. Co-founded Site Engine Inc. in 2008. Founded Unimon Inc. in 2010 and Enison Inc. in 2025, leading development of business systems, NLP, and platform solutions. Currently focuses on product development and AI/DX initiatives leveraging generative AI and large language models (LLMs).