Success Story

Daffodil helps India’s largest automobile manufacturer to develop an AI-driven solution for analysis of part failure images

About the client


The client is India’s largest automobile manufacturer. It is credited with having ushered in the automobile revolution in the country. The company is engaged in the business of manufacturing and sale of passenger vehicles in India. The client has a vast portfolio of 16 car models with over 150 variants. Other activities include facilitation of pre-owned car sales fleet management, car financing.

  • 2 hrs Time Saved Per Defect Analysis
  • 2.5 - 3s Model Prediction Time
  • 88% Accuracy in Failure Mode Detection

Business Situation

The client, a premier automobile manufacturer, faced significant challenges with its complex and labor-intensive motor part failure and fracture analysis process. The existing system, heavily reliant on expert evaluations, resulted in time-consuming, inconsistent, and inefficient outcomes.

Additionally, the company struggled to efficiently respond to sudden or emergency assessment requirements. The categorization of fractures, whether brittle or ductile, and determining the fracture propagation direction, presented considerable difficulties.

Recognizing these challenges, the client engaged Daffodil Software with a clear objective: to leverage artificial intelligence and image analytics to revolutionize their failure and fracture analysis process. The aim was to develop a solution that would not only expedite the analysis, detection, and reporting of failures but also automate and enhance the accuracy of these analyses.

Here are the challenges encountered by the team:

  • The current process at the client’s organization for analyzing fractures in motor parts was meticulous and time-consuming.
  • The process heavily relied on the expertise of professionals.
  • The existing system posed difficulties in accurately categorizing fractures (as brittle or ductile) and determining the direction of fracture propagation.
  • It was not capable of efficiently handling ad-hoc or emergency assessments.

The Solution

The development process presented a multitude of hurdles that demanded strategic navigation and innovative problem-solving. The following elaborates on the specific challenges encountered during the project, emphasizing the complexity of each issue faced by the Daffodil Software team.


  • Developing a user-friendly web portal:

In order to ease the analysis process, a feature-rich, user-friendly web platform / interface / application was designed for our client that allowed analysts to easily upload, analyze and categorize fractures based on microscopic images. ReactJs was used to build the front end, while Python was leveraged for the back-end. The portal/panel allowed its users to view past reports (in PPT & PDF format), review drafts, and accept/reject pending defects; ensuring a seamless, quick and accurate result on part failure.


  • Tackling Limited Data Challenge:

The client’s data was meticulously sorted into various categories, each corresponding to distinct parameters. These categories were based on the type of material (plastic or rubber), the degree of magnification (10x, 20x, 30x, 50x, 100x, 150x, to 200x), the specific grade of the material (PP, PC, PC PBT, PA ABS, etc.), and the type of fracture (brittle or ductile).

  • Utilizing Amazon Web Services’ Ground Truth

One of the primary reasons behind this decision was the strict compliance requirements stipulated by the client. The client had expressed a clear preference for their data not to be transported out of their existing environment. Given these constraints, AWS Ground Truth emerged as the most suitable option. It allowed us to perform the necessary data annotation tasks while adhering to the client’s data security and privacy guidelines.


  • Improvements in Ductile Image Annotation:

Our initial approach to annotating ductile images involved marking features such as cups and cones, flakes, and stress whitening. However, we encountered several challenges with this methodology. From a machine learning perspective, these features were not distinct enough to facilitate accurate annotations. Predicting directions based on these features proved to be a difficult task.

After facing these hurdles, we engaged in discussions with domain experts to find a solution. These conversations led us to redefine our approach to the annotation of ductile images. We decided to introduce new features that would be more distinctive and thus more conducive to machine learning applications.

  • Maintaining Annotation Review and Quality Control:

The introduction of these dots had a direct impact on our project timeline. The need for a re-review meant additional time and resources had to be allocated to ensure the quality and accuracy of the annotations. The process of rectifying these issues and re-reviewing the annotations was time-consuming. Despite these challenges, our team remained committed to maintaining the highest quality standards for the annotations.


  • AWS Sagemaker Notebook Instance Dependencies

To overcome the dependency issues that were hindering the project’s progress, the Daffodil team decided to create a custom UNet model. This model was designed to be compatible with the AWS Sagemaker Notebook Instance and to function effectively despite the dependency challenges.

This custom UNet model not only resolved the dependency issues but also provided the team with a model that was tailored to our specific needs and requirements.

  • Selection of Training Categories:

Firstly, the team had to consider the data count for each category. Categories with a higher volume of data were more likely to provide a robust training base for the model. Secondly, we took into account the features available in the images associated with each category. Categories that displayed more distinctive and easily identifiable features were deemed more suitable for model training.

Additionally, the team had to factor-in the client’s needs and preferences. We had to prioritize those categories that were most commonly encountered by the client in their fracture detection operations.


  • Accurate Detection of Fracture Types:

To overcome limitations of the segmentation model, we decided to introduce classification models into the workflow to overcome this hurdle. These models were aimed at accurately distinguishing between brittle and ductile fractures, thereby improving the overall precision of our project. However, this solution presented another challenge: the timeline.

Training the classification models within this short time frame was a daunting task. Despite this, our team stepped up to the challenge. They dedicated themselves fully to training the classification models and we were able to successfully train the classification models within the limited timeframe.

  • Zeroing on the Right AI Models:

Segmentation & classification models, known for their exceptional feature localization, were considered for this project – as other models, despite their strengths, could not provide the fine-grained localization required for this particular task.

Classification Models

◉ Segmentation Models


  • Direction Marking Logic Adjustment:

Instead of attempting to trace a path from the non-existent origin to the hackles, we decided to reverse our approach. We would start from the hackles and back propagate to find the origin.

This revised logic for direction marking proved to be a more effective strategy. It allowed us to determine the direction in the images, even in the absence of a visible origin.

  • Thorough Documentation Throughout the Project Execution:

During the entire duration of the project, there was a crucial need for us to maintain meticulous documentation. This was in line with the stringent requirements of the Capability Maturity Model Integration (CMMI) Level 5 standards, which demand a high level of process consistency and optimization.

This adherence to CMMI Level 5 standards played a vital role in enhancing the overall quality of our project and ensuring that all processes were optimized and consistent.

Over the course of two months, we produced an impressive total of 43 CMMI Level 5 documents. This vast body of documentation stands as a testament to our commitment to process excellence and quality.


  • Managing Strict Project Timeline:

Our team went above and beyond their regular working schedules to successfully ensure that all tasks were completed within the tight deadline of 2 months. This level of dedication was maintained throughout the project duration, demonstrating the team’s unwavering commitment to the project’s success.


The Impact

Originally intended as a Proof of Concept (POC), the project exceeded expectations & evolved into a full-scale I-Sense project, successfully meeting the stringent CMMI Level 5 compliance requirements. Team Daffodil completed the project within a strict two-month deadline, saving approximately 60 hours per month, equating to 2 hours per defect. The AI model's prediction speed improved significantly, providing accurate failure mode predictions in just 2.5 to 3 seconds. With an 88% accuracy rate in failure mode detection, the tool provided reliable results for informed decision-making. As a result of Daffodil’s intense hardwork, the team was able to secure additional projects to work on other Artificial Intelligence/Machine Learning (AI/ML) use cases for the client.

  • 2 hrs Time Saved Per Defect Analysis
  • 2.5 - 3s Model Prediction Time
  • 88% Accuracy in Failure Mode Detection

Read Related Case Studies

Get in Touch

Sign up for a 30 min no-obligation strategic session with us

Let us understand your business objectives, set up initial milestones, and plan your software project.

At the end of this 30 min session, walk out with:
  • Validation of your project idea/ scope of your project
  • Actionable insights on which technology would suit your requirements
  • Industry specific best practices that can be applied to your project
  • Implementation and engagement plan of action
  • Ballpark estimate and time-frame for development