Experiences
- Reduced the turnaround time of data processing workflow from 24 hours to 2 minutes and eliminated reliance on human supervision by transforming the existing data ingestion process utilizing Scrapy and automated ETL pipelines.
- Interviewed a team of analysts to define KPIs and user statistics, collaborated with the UI/UX team, and led a group of three interns to design user engagement dashboards using Tableau. Driving a 50% increase in productivity of the marketing team.
- Restructured document search engine appropriating key-word based indexing algorithms (RAKE, TF-IDF, and Page-Rank) to improve search speed by 50% and also improved search result relevancy score.
Spyro.ai is a student-led capstone project as part of the MS in Technology Innovation at the University of Washington. Backed by thorough research, Spyro.ai was the only approved student-led capstone project for the year 2020. GIX capstone projects are designed to facilitate the interdisciplinary experience for students via end-to-end challenges, from exploration of ideas to user testing of the finished products. During the bi-quarter project, I have worked on user research, market research, prototyping, user testing, business models, and GTM strategy.
- Collaborated with Ubicomp lab and Seattle children’s to develop a solution to improve Asthma therapy adherence.
- Executed 25 interviews and surveyed ~100 Asthma patients followed by data analysis to identify and prioritize pain points.
- Planned and directed weekly sprints with team to ensure completion of wekly goals, and to mitigate the dependency issues.
- Led a cross-functional team of three to prototype a mobile application using React Native, MongoDB, and Azure to facilitate physiological symptom tracking, medication reminders, and doctor-patient communication for Asthma patients.
- Integrated Fitbit API and Air Quality Index with the prototype to generate personalized Asthma treatment plans.
- Architected web dashboards using Flask, Plotly, and d3.js to visualize and manipulate patient data for medical professionals.
Trestle Labs is a leading assistive technology startup in the Edu-Tech domain. Trestle Lab’s vision is to provide an interactive learning environment for visually impaired communities in vernacular languages and ultimately increase the living standard for them.
- Architected CNN + LSTM based end-to-end OCR system for regional languages using Tensorflow, achieving 92% Acc.
- Pioneered a revenue stream and made the printed content more accessible for the visually impaired by digitizing 10k+ old printed books to accessible digital formats utilizing Vision OCR and GCP. Used in-house OCR systems for regional languages not yet available vis Vision OCR services.
- Directed usability tests on an in-house mobile application (KIBO) to understand educational requirements of visually impaired and designed features of audio highlighting, content summarization, and auto-linking content with additional online resources.
- Designed a responsive website for Trestle Labs to collect customer feedbacks and promote the service.
DISQ is a social entrepreneurship and innovation facility by Tata Consultancy Services(TCS, India).
- One of the 40 students selected across the geography of India for the Social Innovation fellowship. The fellowship experience included learning opportunities from experts in the field as well as applying the learnings to real-world challenges.
- Worked to solve global challenges from varied domains, such as enhancing the learning experience for the visually impaired, to increasing the shelf life of perishable food in remote areas where electricity is yet to come.
- Together with sharpening technical skills, also worked on user experience, financial and marketing aspect of the idea to create a sustainable solution.
- Pitched the solutions to many potential investors and continuously improved on the business aspects as well as to gain communication and leadership skills in a corporate environment.
Projects
- Repurposed BERT to analyze the sentiment of Tweets and FinViz using Tweepy and Scrapy while eliminating spam content.
- Conducted a time-series analysis to engineer features based on financial market anomalies such as the day of the week effect.
- Programmed a market prediction function using random forest using an ensemble of sentiment analysis, stock indicators, and market anomalies on AWS Lambda and connected the prediction function with Alpaca API to automate the trading process.
- Research on privacy-preserving real-time cough detection
- Conducted market research to verify the need for privacy-preserving and ubiquitous respiratory symptom tracking devices.
- Constructed an end-to-end audio classification model using PyTorch, eliminating the need for acoustic data preprocessing.
- Optimized and deployed audio classification model on the TWS earphones utilizing quantization and CMSIS-NN implementation to monitor physiological respiratory indicators such as Cough, Sneezing, and Wheezing with 97% accuracy.
- Submitted the findings at Ubicomp comference - August 2021.
- Created a Decision Tree classifier to detect bone fractures by inducing acoustic resonance in bones using sound waves.
- Designed wire-frames for mobile application UI, conducted usability tests, and created animated video tutorials.
- Built a platform-agnostic mobile app using Flutter and connected it with fracture detection API deployed in Azure.
- Recognized by IEEE-GHTC-2020 for developing a 92% accurate, low-cost and non-invasive fracture detection system inteded to use in rural low resource areas in 3rd world nations.
- Implemented transfer learning on the state of the art object detection algorithms including DETR, Efficient-DET, Faster RCNN, RetinaNet and YOLOv5 to detect wheat heads from crop images. Achieved highest MAP score of 0.65.
- Used NMS(Non-Maximal Suppression) to make ensembles from above-described models and improved score to 0.69.
- Replaced NMS by implementing WBF(Weighted Box Fusion), which increased the score to 0.71.
- At last implemented Pseudo-labelling, to deal with a smaller dataset, which resulted in a final score of 0.725. (top 10%)
- Used Tweepy to extract the 50K Tweets, extracted Tweet sentiment and context using NLTK, as well as extracted graph based features of Twitter users by performing Clique and Community analysis(NetworkX). Achieved 75% accuracy in tweet virality prediction using Random-Forest algorithm(sklearn).
- Engineered statistical evidence for the time-dependent recurring market anomalies(day of the week, the month of the year, etc.).
- Developed a 1D CNN based approach to detect the binary market movement for individual stocks based on historical data.
- Achieved 7% better accuracy utilizing ensemble of traditional stock indicators and features based on market anomalies as compared to solely using traditional indicators.
Publications
IEEE-GHTC-2020
International Conference on Computing, Communications, and Cyber-Security (IC4S 2019)