TeamKART: Official FSAE Team of IIT KGP

Website

TeamKART is a group of students from Indian Institute of Technology Kharagpur who like to explore the practical side of automotive engineering. TeamKART is officially a sponsored research project under the Sponsored Research and Industrial Consultancy Cell (SRIC) and the Department of Mechanical Engineering, IIT Kharagpur.

I contributed to the project K5 which has been one of the most successful endeavours of the team so far. As a member of the Powertrain subsystem, I collaborated in the design of the following components for the vehicle:

• 3D printed Intake manifold
  FSAE rules dictate that the intake system should consist of a restrictor of 20 mm dia. in order to limit power output of the engine. Due to the presence of a restrictor, a significant pressure drop is generated and this drop is recovered with a diffuser, after which the mass of air enters into the plenum, which serves as a reservoir of air, almost at atmospheric pressure. Following the plenum, the last component is the runner whose length is optimised to generate max engine power at a given RPM. We use KTM DUKE 390 which is a 4 stroke single cylinder engine. The entire design is initially simulated on Ricardo WAVE and based on the results obtained, the steady-state Ansys Fluent Analysis is performed. This process is iterated until an acceptable engine performance and fluid behavior is obtained.
• Fuel Tank
  Our fuel tank is made using industrial grade Aluminium. This ensures significant weight reduction compared to using steel. The key feature of the design is to ensure minimum pressure drop on the fuel pump so as to prevent any disturbance to the engine's fuel supply. We introduce baffles to serve this purpose. Lastly, the mounting of the fuel tank should be such that it can sustain the longitudinal and lateral accelerations that the vehicle can achieve, keeping the flexibility in mind. The total capacity of the fuel tank to sustain the Endurance Event in FSAE competitions is estimated by simulations using Optimum Lap developed by OptimumG.
• Drivetrain arrangement
  The drivetrain package is composed of a chain-sprocket drive, Drexler Limited Slip Differential and driveshafts for the rear wheels. We use Optimum Lap developed by OptimumG as a simulation tool to calculate the optimum Final Drive Ratio (FDR) to achieve the desired lap times in the dynamic events in FSAE competitions, using the engine and vehicle's parameters. After calculating the FDR, we design the sprocket and perform steady-state static and fatigue analysis on the same to ensure its reliabilty. The driveshafts are designed considering the required lengths, diameters of the wheel hubs and CV joints. It is ensured that the 2 driveshafts have equal torsional stiffness.
• Downflow Radiator
  We designed the radiator considering the heat required to be dissipated and packaging constraints. Our primary goal was to calculate the required convective area satisfying the constraints. Using the LMTD method to perform the calculations, we obtained the required dimensions and then got it manufactured. The airflow was enhanced using 2 axial fans selected to ensure the required mass flow rate of air for efficient heat rejection.

With respect to the team's Electric vehicle design, I contributed to the following:

• Motor selection and transmission system
  We shortlisted a variety of motors based on cost, availablity in India, controllability, etc. Following this, we performed multiple simulations on Optimum Lap to select the motor to achieve the required lap time. Planetary gear arrangement coupled with chain-sprocket drive was the transmission system finalized to satisfy the torque requirements of the vehicle. Steady-state static and fatigue analysis were performed to finalize the design and material selection.
• Cooling system
  The selected motor was air-cooled by design. We performed thermal analysis to ensure its cooling performance. The motor controller was cooled by forced convection using heat sinks and axial fans. All the simulations were performed on Ansys ICEPAK.

For our current project K6, we are introducing multiple design changes such as aerodynamic package, 10" wheels, telemetry system compared to previous vehicles in order to enhance the vehicle's performance in the dynamic events. One of such enhancements is the reactive muffler. Based on our literature review, we finalised 2 variants considering the required insertion loss of 25 dB, minimum backpressure and weight. The fluid analysis was performed on the Ansys Fluent to estimate the backpressure and understand the fluid motion. Acoustic simulations on COMSOL Multiphysics to estimate the transmission loss and compare the designs.

References

• Design and fabrication of intake manifold for formula SAE (Society of Automotive Engineers) race car by Sylvie Doré, Patrice Lavallée
• FSAE Electric Vehicle Cooling System Design by Jeff LaMarre
• Design principles for an automotive muffler by Vinod Sherekar and P. R. Dhamangaonkar
• Design and fabrication of reactive muffler by G. Gnanendhar Reddy and N. Prakash

Close

Mathematical modeling of Li-ion batteries with focus on Silicon anode particles

Supervisor: Dr. Jeevanjyoti Chakraborty, Mechanical Engineering Department, IIT KGP

A crystalline silicon anode has a theoretical specific capacity of 3600 mAh/g, approximately ten times that of commonly used graphite anodes. This makes it a suitable candidate to be used as an electrode in Li-ion batteries in order to maximise battery's capacity and minimize its weight simultaneously. However, during the charging and discharging of the Lithium Silicon battery, the lattice distance between silicon atoms multiplies as it accommodates lithium ions (lithiation), reaching 320% of the original volume. Due to this massive expansion, stresses of large magnitudes are generated which ultimately leads to the fracture of the silicon material.

This project is aimed at understanding the behaviour of Silicon in these batteries, identify the stress generation and propose a solution to this problem in the long run. The mathematical modelling approach was selected because of the involvement of Partial Differential Equations (PDE) and numerical analysis. In the first stage of the project, I studied the stress generation and diffusion of Li+ ions during potentiostatic (constant voltage operation) and galvanostatic (constant current operation) by solving the PDEs with appropriate boundary conditions. The stress due to diffusion (lithiation) was considered.

The next stage of the project was to formulate a bidirectional relationship between stress and Li+ ion concentration over time. This was considered to be crucial because of the fact that as time progresses, stress increases and hence can have an effect on the concentration which in turn would affect the cell chemistry. The concept of chemical potential was used to establish a connection between the stress and concentration. I solved the involved PDEs using Python and modelled them using COMSOL Multiphysics for visualisation.

In the final phase of my contribution, we moved on to identify crack development behaviour in the Silicon material due to the generated stresses. The amorphous lithiated silicon formed around the crystalline Silicon anode has different regions with elastic and plastic properties. Due to this, the mechanical equilibrium and diffusion equations have different boundary conditions.

References

• Evolution of stress within a spherical insertion electrode particle under potentiostatic and galvanostatic operation
• Fracture Modeling of Lithium-Silicon Battery Based on Variable Elastic Moduli
• Stress evolution with concentration-dependent compositional expansion in a silicon lithium-ion battery anode particle

Close

Fault Tolerant Control System for Electric Vehicles

Supervisor: Dr. Somnath Sengupta, Advanced Technology Development Center, IIT KGP

To be updated soon. Stay tuned!

Close

Self Driving Cars Specialization

Coursera

Github Repository | [Certificate]

This is a project-based specialization comprising of 4 courses and individual project at the end of each course. These courses provide extensive understanding about self-driving cars including the technical aspects, current advancements and safety standards.

The 4 courses are as follows:-

• Introduction
  This course gives a introduction about self driving cars, understanding about ODD and environment and fundamental concepts regarding longitudinal and lateral vehicle dynamics, equation based modelling approach and control strategies. For the final project, I had to implement vehicle control strategies to determine the optimum throttle, steering and brake values required for the vehicle to traverse in a racetrack environment in the open-source CARLA Simulator. I implemented PID, Linear Quadratic Regulator (LQR) and Model Predictive Control (MPC) for longitudinal control alongside Stanley Steering Control and PID for lateral control.



• State Estimation and Localisation
  In this course, the concepts of state estimation using Kalman Filters (Extended, Unscented, Linear), localization and sensor fusion were introduced. I had to deploy an Error State Extended Kalman Filter for state estimation using IMU, GPS and Lidar datasets from CARLA simulator, and at the same time, keep track of the vehicle's trajectory (3D) in real-time. The consistency of the algorithm was checked via introducing situations like absence of sensor data and faulty sensor data.
• Visual Perception
  This course contained the description related to the main perception tasks in autonomous driving, object detection and computer vision methods. Also, I worked with camera calibration model, and semantic segmentation. For the final project, I developed algorithms for bounding boxes around objects detected by the Convolutional Neural Network model and finally perform drivable space estimation.
• Motion Planning
  From this course, I learnt about finding shortest path over a graph or road network using Breadth First and Depth First search, Dijkstra's and the A* search algorithm. Also, I used finite state machines to select safe behaviors to execute, and design optimal, smooth paths and velocity profiles to navigate safely around obstacles while obeying traffic laws. For the final project, I implemented a hierarchial motion planner to navigate through a variety of scenarios in the CARLA simulator including parked vehicle, stop signal, and following a lead vehicle.

Close

Deep Reinforcement Learning for Autonomous Vehicles

Objective: To use RL framework for safe traversal of a self-driving car in CARLA simulator environment.

Github Repository

Reinforcement learning (RL) is an area of machine learning concerned with how intelligent agents ought to take actions in an environment in order to maximize the notion of cumulative reward. Reinforcement learning is one of three basic machine learning paradigms, alongside supervised learning and unsupervised learning. RL has very promising applications in various aspects of autonomous driving.

In this project, I used the Deep Q-Learning algorithm with Temporal difference learning update rule and dueling deep learning framework with value and advantage functions to compute the action values. I implemented a 3-layer CNN with the image from the onboard camera as the state input. The output of this layer is flattened and then divided into two outputs - a value function and an advantage function, which are then used to compute the action values. The action space is composed of discrete tuples of the throttle, brake, and steering values. The action values obtained from the neural network are mapped to discrete values, using the euclidean norm. These actions were then sent to the CARLA controller. Finally, the real-time reward is computed using the data from the wheel odometry, lane invasion, and collision sensors. Needless to say, I used separate training and target networks. For the experience relay, I used the Sum tree data structure for having a prioritized experience resampling.

In the intial stages of the simulation, the vehicle's steering response was unstable because of discrete values and low training. With time, this behaviour was minimised and the lateral stability was obtained. I ran the simulation for about 45 mins which covered almost 60% of the pre-planned trajectory. To understand the agent's performance on the vehicle control quantitatively, I used Stanley Steering control to compute the steering values at every time instant required to traverse the path.

References

• Reinforcement Learning with Tensorflow by Sayon Dutta
• CARMA: A Deep Reinforcement Learning Approach to Autonomous Driving by Matt Vitelli and Aran Nayebi
• Behavioral Planning for Automated Vehicles using Deep Reinforcement Learning by Meha Kaushik
• A Reinforcement Learning Based Approach for Automated Lane Change Maneuvers by Pin Wang, Ching-Yao Chan and Arnaud de La Fortelle

Close

Traffic Sign Classification using CNN

Github Repository

The convolutional neural network (CNN) is a class of deep learning neural networks. CNNs represent a huge breakthrough in image recognition. They’re most commonly used to analyze visual imagery and are frequently working behind the scenes in image classification. They can be found at the core of everything from Facebook’s photo tagging to self-driving cars. Image classification is the process of taking an input (like a picture) and outputting a class (like “cat”) or a probability that the input is a particular class (“there’s a 90% probability that this input is a cat”).
A typical CNN consists of - Convolution layers, activation functions, pooling layers, fully connected layers.

In this project, I used a readily available dataset on Kaggle and deployed a Convolutional Neural Network model to act as a classifier for 43 different classes of traffic signs.
My neural network architecture: Input image -> ( Conv -> Max Pooling ) -> ( Conv -> Max Pooling ) ->( Conv -> Max Pooling ) -> Dense -> Dense -> Dense Layer (softmax function).

After the first training and testing session, the testing accuracy wasn't satisfactory and the network seemed to overfit the given dataset. To rectify this, I changed the architecture by introducing a Dropout layer with 40% as chosen value and a fresh convolution + pooling pair. This led to a significant change in the testing accuracy and the problem of overfitting was rectified. I tested the framework over about 12,000 images and obtained an accuracy of 94%. For the complete project, Python programming language was used with the mentioned libraries - Numpy, Tensorflow, Keras, Matplotlib and OpenCV.

Close

Inter IIT Tech Meet 9.0

Event: Bosch's Electric Vehicle Simulation

Presentation video (starts from 47:00) | Problem Statement | [Certificate]

This was a virtual event hosted by IIT Guwahati, in March 2021. I was a part of the 8-member team for the event, "Bosch's Electric Vehicle Simulation". Our solutions for PMSM motor design and control, battery dimensioning and powertrain architecture were highly appreciated by the judges.

1. Secured the Gold medal in the event, for the proposed solution.
2. Proposed the powertrain architecture comprising a PMSM motor (self-designed), ABS, VSC, etc.
3. Determined the motor’s requirements and battery pack’s capacity based on selected target performance parameters

Close

FSEV Concept Challenge 2020

Virtual event hosted by Curiosum Tech

[Certificate]

This event was held remotely wherein we submitted the presentation and several documents to support our design and management reports as well as procurement strategies for the production of a Formula-style electric vehicle. Improving upon our previous performance, we secured:-

1. 5th Place among 37 teams internationally.
2. 2nd Place in EV presentation 4th Place in Powertrain Package Design
3. 4th Place in Powertrain Package Design
4. Notable mention - Fresher Team Effort: Engineering Design

Close

Formula Bharat 2020

Hosted by Curiosum Tech.

Formula Bharat 2020 Promo video | [Certificate]

I, as a part of TeamKART, participated in Formula Bharat 2020, held in January 2020 at the Kari Motor Speedway, Coimbatore, Chennai. India. TeamKART achieved the 'one-year-one-car' goal, or in other words, designing, manufacturing, and testing the vehicle in a year. This was one of the most successful endeavors of the team so far. We secured:

1. Overall 10th place among 51 participant teams in the Combustion category.
2. 2nd Place in Business Plan Presentation
3. 6th Place in Engineering Design Event
4. 15th Place in Cost and Manufacturing Event

Close

Flipkart GRiD 2.0

Hosted on Dare2Compete

[Certificate]

I was a part of the 4-member team which participated in the Robotics Challenge with the problem statement - "Autonomous Indoor Drone". We submitted the solutions for monocular vision-based navigation, object detection, autonomous control and planar obstacle avoidance. The selected flight controller was ArduPilot and the selected microcontroller (for other tasks) was Raspberry Pi 4. The camera, barometric pressure sensor and obstacle avoidance sensor were used.

1. One of the top 20 teams to clear Round 1 of the challenge.
2. Successfully cleared levels 2/4 of the competition.

Close

Summer Intern

KPIT Technologies Ltd, Pune, India

Certificate

My primary responsibility during the internship was to explore the APIs of AI compilers - TVM, Glow and MATLAB and leverage them to generate hardware-specific and optimized executable bundles which can be easily deployed on hardware backends. Deep learning models are significantly heavy when it comes to memory usage and time required to compute the results. This makes their deployment on edge devices, having lesser RAM and ROM, tedious. The existing DL frameworks namely, TensorFlow, PyTorch and ONNX provide support for training and deployment on CPUs and GPUs. However, when it comes to edge devices with low power capabilities, the optimization of these models becomes a challenging task.

In the first part of the internship, I leveraged the AI compilers to optimize well-known pretrained DL models in existing frameworks for LLVM and GPU devices. The customization options vary with respect to the compilers and each of them had different optimization schemes.
The models were selected or created so as to ensure that different types of architectures are considered. Then we decided the performance metrics which were then used to observe the relative performance of these compilers and frameworks. These metrics were essentially the representatives of the memory consumption and inference time of the selected models. Finally, we observed the relative behaviour of these compilers in terms of the performance metrics, their compatibility and efficiency to optimize the different model architectures and hardware-specific deployment capabilities they offered. The comparison performed was qualitative as well as quantitative so as to provide a better method of decision-making as per the user's application.
The codes were based on C++ and Python, varying slightly with respect to the compilers' and frameworks' API.

Close

UG Research Assistant

Centre for Automotive Research, OSU

Certificate

The project dealt with conducting the performance evaluation of a medium fidelity Driver-in-the-Loop Simulator, present in the Driver Dynamics Lab at OSU. In the initial phase of the project, I reviewed the literature for Objective Motion Cueing Test (OMCT). It is applied by using a series of sinusoidal signals from 0.1 rad/s (0.0159 Hz) to 15.849 rad/s (2.515 Hz) - to each of the input channels of the Motion Cueing Algorithm (MCA), in order to measure the output amplitude and phase distortion of each of these channels. OMCT is a frequency-based method.

Based on the specifications of the simulator, we finalized the criteria and testing methodology using input signals generated using Simulink which are then applied to the platform using the simulator's interface. Finally, in order to capture and estimate the velocities and accelerations in 3D, we used linear variable differential transformers (LVDTs). The results of this experiment helped us in identifying the fidelity regions of the simulator with respect to its motion about different axes, using the modulus and phase plotted against the frequency of the signal. To get rid of the noise in sensor readings, we used the Butterworth filter.

In the next phase of the project, we looked into the transformation and filtering of the data collected from an inertial measurement unit (IMU) to be deployed on the simulator to capture its acceleration in different directions. The transformation was required so as to estimate the acceleration at a different point compared to the IMU's location. We collected the required data for 2 locations on the ego vehicle from CarSim for different maneuvers. In order to mimic the sensor's behaviour, we introduced random noise in the signals and deployed different filtering techniques to remove the same as much as possible. Using physics-based transformation for rigid body dynamics, we obtained the acceleration at the required coordinate in 3D and validated this against that obtained from CarSim. Finally, we implemented a Gaussian filter of configurable window length and the obtained errors were within acceptable limits.

The codes were based on MATLAB.

References

• Evaluation of Motion Cueing Algorithms for a Limited Motion Platform Driver-in-Loop Simulator by Rubanraj Sekar
• Objective Motion Cueing Test for Driving Simulators by M. Fischer et al.
• On the Objective Evaluation of Motion Cueing in Vehicle Simulations by Sergio Casas-Yrurzum et al.
• Relative motion of points on a body

Close

Winter Intern

ATI Motors, Bengaluru, India

Certificate

As a part of this internship, I was required to explore state-of-the-art SLAM algorithms and select the one feasible and favourable that can be applied to a variety of existing datasets that the Autonomy team possessed. During the entire course of this internship, primarily, I focused on OctoMap to visualise 3D probabilistic maps in an efficient manner and Particle Filter SLAM or Gmapping SLAM.

In the first part of the internship, I studied about OctoMap and experimented on a couple of open-source datasets available on University of Freiburg's website. Next, I explored SLAM algorithms like IMLS SLAM, LOAM SLAM, EKF SLAM and Particle Filter SLAM. Based on the extent and efficiency of the algorithms, we finalized the Particle Filter SLAM for implementation. The existing algorithm that the team used was scan-matching with a predefined grid. The team was in the possession of multiple lidar and wheel odometry datasets collected from their robot subjected to a variety of environments.
After the implementation of the algorithm, we checked for its consistency using the datasets. We got satisfactory results with the initial algorithm. We tried out a couple of resampling algorithms and as expected, the quality of the generated map improved significantly. The quality of the map was calculated using the entropy function. Also, I studied the effect of variation in parameters like number of particles, motion noise, distance filters for lidar, etc. on the map quality.
The entire algorithm was implemented in Python using the Numpy, Matplotlib and Numba (for reducing time complexity) libraries.

References

• Prof. Dr. Cyrill Stachniss’ lectures on SLAM algorithms
• On Resampling Algorithms for Particle Filters by Jeroen D. Hol, Thomas B. Schön, Fredrik Gustafsson
• A Probabilistic Online Mapping Algorithm for Teams of Mobile Robots by Sebastian Thrun

Close

Research Intern

ARMS Lab, IIT Bombay, India

Project title: Applications of Machine Learning Methods for intersection and lane-following for autonomous vehicles

Supervisors: Prof. Leena Vachhani | Prof. Arpita Sinha

Initially, I implemented a Finite State Machine consisting 4 states - Decelerate to stop, Stop, Follow lead vehicle, and Track speed with required transitions. This was done to understand the agent's behavior due to the brute force approach and hence can be used for comparing with results from the ML model OR this machine can be enhanced with Machine Learning tools. I was asked to define various situations including intersection handling and lane-following in the presence/absence of obstacles. This was tested in the CARLA simulator.

In the next phase of the project, I devised an automatic controller in Python that was used to roam around in the CARLA environment and collect datasets. These datasets comprised of images from the onboard camera sensor and real-time client measurements consisting of control commands, distance to nearby obstacles, etc. The Town-01 of the CARLA simulator was used. The ego vehicle as well as the other agents were modified to violate traffic light regulations. The controller incorporated collision avoidance which was activated/deactivated to collect a variety of data points favourable for the efficient training of the neural network.

Towards the end of the project, I had to define the complete problem using a Markov Decision Process representation with states, actions, rewards and subsequent states. The reward was evaluated as a discounted sum for a particular episode. But, these values might not be consistent due to variation in the reward accumulation and randomisation in the agent’s action selection behaviour. Hence, I had to normalize the rewards so as to avoid incorrect computations. The normalization was performed using the mean and standard deviation of the rewards.

Close

Research and Development Intern

Kanan Park, Pune, India

Certificate

In the beginning of the internship, I was working on the autonomous landing of the drone using Raspberry Pi's image processing and Interrupt Service Routine. The autonomous control of the drone was achieved using GPS waypoints. The waypoints were preset in Ardupilot's Mission Planner. The image processing was performed to detect ArUco marker placed on the ground and then the landing sequence was initiated with a delay. This was tested with Hardware-in-the-loop simulation.
In order to increase the robustness of the drone's landing sequence, I suggested the usage of Reinforcement Learning. The Deep Q-learning approach was finalised with suitable update algorithm and reward function. I formulated the conceptual model for this idea.

Next, I was assigned to collaborate on the motorized valve control for remote water flow system in farms. GSM signal from the farmer's mobile was used to actuate the motor. I worked on the code for the Arduino end which dealt with reception of signal and actuation of the motor. Apart from this, I designed the mechanical components using gears and support mechanisms.

Lastly, I designed a multi blade grass cutter to be deployed on the remote-controlled rover. Also, I performed the Finite Element Analysis on Ansys to study the deformations of the mechanism.

Close

Close

Close

Close

Close

Close

Close

Close

Close

Close

Mental Health - It is really important??

Link to Blog

I published this blog on July 22nd, 2020 highlighting the importance of mental health and how people, in general, tend to neglect this. This blog covers preliminary info regarding mental health, myths, and beliefs in India about this, statistical analysis, and an exemplary video to explain the current situation of students. P.S Don't get confused by the title. It is meant to be in a sarcastic sense. :) I concluded the blog with a few words of my own. ;)

Leave a comment and follow me on Medium, if you wish!!

Close

Kedarkantha Trek: Freezing weather, Camping & the Himalayan Summit

Link to Blog

I published this blog on January 1st, 2022 recollecting my first-every trekking experience to one of the most popular peaks of India - The Kedarkantha (12,500 ft). This blog encompasses my detailed experience starting from day-0 including the food, living conditions, and more. I have added some pro-tips based on my knowledge and encounters.

Leave a comment and follow me on Medium, if you wish!!

Close

Can we do everything?

Link to Blog

I published this blog on January 13th, 2022 in an attempt to answer the ubiquitous question (for myself) - can we really do everything that we want to, in our lives? This blog addresses how motivational speakers make it sound so easy whereas in reality, this question deserves a better answer.

Leave a comment and follow me on Medium, if you wish!!

Close

Safety Measures underlying the thrills and spills of F1

Link to Blog

I published this on December 2nd, 2020 marked by the fireball crash of Haas F1 Driver Romain Grosjean. This blog pinpoints the safety measures that have been incorporated in F1 over the years for the drivers along with focusing on the importance of safety, especially when it comes to driving.

Co-author - Swarnim Raj: LinkedIn | Facebook

Leave a comment and follow me on Medium, if you wish!!

Close

CDC Internship Diary -KPIT Technologies Ltd.

Link to Blog

I published this on January 16th, 2021 highlighting the key aspects of the CDC internship procedure, my experience of the interviews and tests involved, some facts and advice about the entire process.

Leave a comment and follow me on Medium, if you wish!!

Close

A Holistic Guide to Research Internships - Domestic and International

Link to Blog

I published this blog on February 23rd, 2021 which can serve as a guide to the students interested in applying for research internships. I have added my experience as well at the end.

Leave a comment and follow me on Medium, if you wish!!

Close

Efficient Multitasking - Is it an oxymoron?

Link to Blog

I published this blog on July 6th, 2021 to highlight how multitasking, as is, in real life, is a proven myth. There's a substantial number of research studies conducted to prove that human beings are not designed to multitask. In the blog, I have presented conclusive results from research conducted by neuroscientists and psychologists, why can't we multitask and methods to avoid being in such a scenario.

Leave a comment and follow me on Medium, if you wish!!

Close

CDC - Core / Non-core / FT - Oh, what the hell!

Link to Blog

During this time of the year, students, in IITs, are perplexed about on-campus internship drives and research internships as well. In this blog, published on 9th July 2021, I have attempted to alleviate their stress levels by providing an overview of how the upcoming couple of months are going to look like and what are the different ways, they can secure an internship offer.

Leave a comment and follow me on Medium, if you wish!!

Close