Aditi, Khandelwal

Internship Description

Aditi Khandelwal, ’20SEAS, interned with Wavelength, a firm that helps existing buildings reduce energy wastage to comply with New York City’s most ambitious energy requirement policies and reduce their carbon footprint and meet sustainability clauses, since only two percent of New York City’s structures are responsible for 48 percent of the city’s energy use. Aditi conducted data analysis using open source data and derived insights for Wavelength on their target buildings to assist them with reducing their carbon footprint.

This summer, I interned as a computer vision data scientist with Wavelength Lighting, a startup in the sustainable lighting and efficiency space. The goal was to build a computer vision app that identifies bulb fixtures when an image is uploaded on to it. This app would help businesses and Wavelength’s clients detect which type of bulb fixture damaged or needed to replace by clicking a picture and getting the result as the ‘bulb type’ on the app. There was no preexisting data science team, therefore I handled the project independently end-to-end.

There were three stages to my project: data collection; model; deploy. My roles and responsibilities included the following:
- Brainstorm, develop, and structure the timeline of the product
- Scrape images of different type of fixtures from Google Images to create a dataset to work with
- Clean data by removing duplicates, incorrect labels, and structuring the data for input into a CNN
- Build a multiclass CNN model to identify different kind of fixtures by tuning hyperparameters and transfer learning by measuring precision and recall of the model
- Deploy the model using Flask, host it on the web as an open source application for users to test
- Integrate the app to the Wavelength website for users to use
- Build, monitor, and maintain and improve the model as and when we obtain real time data
- Help build a sales scraper for the sales team, which scrapes contact details of prospective clients from LinkedIn so that they can reach out to them without looking through and getting their numbers

I worked directly with the CEO of the company. We met once a week to determine the scope of the project, the pipeline, and its outcome. I handled the project end-to-end independently due to the absence of a dedicated data science team, delivering a product that was entirely new and fascinating to the Wavelength team. This role was an amalgamation of a computer vision data scientist as well as a product manager. My knowledge from business courses, as well as technical courses in the world of sustainable energy, helped me build a product from scratch using computer vision. I was able to use the skills learned at Columbia in a professional setting, which was accompanied with its own set of challenges. I learned how to communicate with supervisors from a non-technical background, how to manage expectations, and how to build a product journey from its birth stage to real-time production.

Thousands of dollars are being invested in the sustainable lighting and energy. I was excited to work with a company that works in energy consumption, clean energy, and efficiency, and to be associated with a firm that aim towards making New York City more sustainable in the future. I am happy to be part of this change and work with some of the city's most notable buildings and locations. My company aimed at producing an inhouse bulb classifier that helps people at both a household and corporate level to solve the simple problem of bulb detection. Often, individuals are not able to understand what bulb type they are using, or what bulb they need to replace because most of the time these bulbs are not practically labelled. Therefore, the need of a bulb classifier arises.  

A regular day during the summer would begin with 30-minute daily morning check-in with my manager and another intern. Both of us gave a summary of the previous day’s work, plan for the current day, and any issues or suggestions we had. I would also have brainstorming sessions with the CEO once a week to give my product more direction and find use cases.

I built a dataset, modeled the multiclass classification with an accuracy of 89 percent, and hosted it on the web for users to upload bulb images and get their prediction. Wavelength aims to incorporate this app into their website. The app allows people to get to know and recognize their bulbs, saving them time, effort, and money.

Overall, the internship was a rigorous learning experience as there was no one to go to with technical errors or issues. The fact that I had to debug, struggle, and build the product end-to-end from scratch bought an immense feeling of satisfaction and learning. I was given the freedom to explore different techniques and give the product a direction. I practically applied the knowledge I gained at Columbia — particularly through courses like artificial intelligence, machine learning, data analytics, and tools for analytics. I also applied my learnings from online summer courses to finesse the hyperparameters and improve the accuracy of the CNN. I was exposed to how to work in a non-technical environment and effectively communicate. My team appreciated my explanation of the technical aspects of CNNs and how neural nets work. I honed my soft skills by interacting with the team and CEO daily and put my thoughts and ideas forth towards the product. A key takeaway is to believe in yourself and explore the options and resources available to you. I built and delivered a product, despite not knowing if I would be able to deliver it on the outset. I would like to thank Mr. Michael Hennesy (CEO), my project manager Sara Annis, and the IEOR department for giving me the opportunity to develop this product from the ground up, and which I get to call my own.

The App is open source and can be accessed by anyone at https://bulbdetection.herokuapp.com/templates/.

The entire project code can be found at: https://github.com/aditi310896/wavelengthmodel