At Ninjacart, we are working on various projects that range from core computer-vision to reinforcement learning.
- Understanding the quality of Vegetables and Fruits through computer vision,
- Solving vendor-related problems (from matching to fraud detection),
- Inventory management with optimized efficiency in logistic operations,
- Supply-demand-price forecasting,
- High dimensional satellite imagery analysis, and
- Building ML-Ops tools and pipelines to create the right data science approach, form the core of our Research efforts.
A few of the technical research modules of these projects that we are currently working on are as follows:
Semi-Supervised Instance aware segmentation: Computer Vision
We have used a modified version of Deep MARC to create a segmentation approach for instance-level segmentation. The amount of labelled images that we have to use through this approach is drastically lower, compared to any existing Instance level models. We’re looking forward to working on lightweight models which could achieve similar results.
Monocular Depth Estimation: Computer Vision
Using a dynamic-bin strategy through encoder-decoder-transformer architecture is something that is working for us. We have followed innovative approaches to create labelled data for single image depth maps. We have ideas to work on this further to make the depth estimation more accurate and faster.
Finding Image Orientation and extracting 3D Model from 2D Image: Computer Vision
The primary problem statement is to understand orientation around the object-of-interest by extracting key landmarks, then using this orientation to extract a 3D model of the object. We have been working on a few approaches, using multiple images through different capture points.
We are planning to work on using stereo cameras for the same. This is one of the primary problem statements, on which we are excited to work in the coming days.
Deep learning on the Edge.
This is another primary area of focus at Ninjacart. We are interested in lightweight models with inference speed as the primary metric in deploying models. In this problem statement, the edge can be an android/iOS/embedded device.
Graph Neural Network modelling to understand the relationship between vendors.
The idea is to understand the relationship between different vendors for better matching. We are also trying to establish the health/quality of a vendor by taking into account the breadth and quality of his network. This metric can aid in creating a differential treatment for vendors consistently indulging in good vs fraudulent behaviour.
Usage of RL to optimize Logistics related problem statements.
This is one of the ambitious goals we are working towards.
We have created a digital twin of the entire supply chain logistics, and the idea is to use RL to optimize different blocks in this digital twin atmosphere and replicate these changes in the real supply chain. We are open to more ideas on how to take this forward.
Supply, Demand, Churn, and other predictive forecasting models.
A lot of current Ninjacart business revolves around forecasting models. Be it price estimation, churn prediction, supply-demand forecasting, and so on. The work on this is continuously evolving with respect to business. We are looking out for more ideas in this direction on how we can effectively solve these problem statements in a better and more optimized way.
NinJutsu is an internal tool that we are planning to open-source, which can manage your development, deployment, and monitoring issues. We will try to share more reading material around this as we make further progress.
If you have relevant experience in the fields of Data Science and ML-ops and want to be a part of our team, do send your CV directly to email@example.com with DATA SCIENCE as the subject line.