Aug 2021 – May 2022 · 9 mos · India

• Unicorn Startup. Parent - DailyHunt (News Aggregator) and Josh (Short Video)
• Worked on de-duplicating news articles based on various signals such as document vectors, time difference, publishers and named entities across various languages - English and Indic.
• Using FAISS for semantic searching of vectors based on document vectors across millions of articles across relevant time periods.
• The document vectors and named entities were used in subsequent recommendation systems based on people and content groups and entity linking through named entity graphs.
• Deployed the solution with low latency and high throughput configuration in a highly scalable architecture.
• Increased the time spent on the app and unique user clicking more times on articles by a large amount.
• Lead a team of 4 (2 FTE + 2 interns) across different projects as a track lead.

Jun 2019 – Jul 2021 · 2 yrs 1 mo · Delhi, India

• Unicorn Startup. Parent - Hike Sticker Chat (Messaging), Rush (Gaming), Vibe (Social Media)
• Project 1 - HikeMoji Personalization
• Overview: Created a 2D and 3D virtual avatar of a person from their photo/selfie.
• Facial Feature Identification - Identified (classified) hair, eyes, nose, lips, earrings, outfit of a person based on their picture using CNN, DLIB, Logistic and Linear regression techniques aggregated in novel ways using Pytorch, Fastai and Tensorflow backend.
• Face Shape Analysis - Geometric self-supervised way of matching person's face-shape to the avatar.
• Colour Prediction - Identified the colour of the ROI based on clustering approach and heuristics on top.
• Tagging pipeline - Designed the pipeline for tagging of images created feedback loop to get highly accurate data and deployed APIs in FLASK.
• Cloth variation - Derived 'infinite' user defined clothes textures through a photo.
• Serving - Deployed and served the models on a Python backend with low latencies (under 2 seconds) and high throughput.
• Result: Achieved high (>90%) accuracies in various tasks, high customer satisfaction rates, reduced funnel drops in creation flow.
• Project 2 - Hikemoji Animation
• Overview: Identified the various actions of a person in their video and replicated the same on their virtual 2D character.
• Facial Motion Identification - Identified actions like pose, eye-blinking, eyebrow movement, eyeball-movement using deep learning techniques and facial landmark extraction aggregated with heuristics.
• Lip Sync - Trained a decision tree classifier on live audio from the microphone for phoneme extraction in unity in real time.
• Mobile Facial Landmark Detection - Tried building face bounding box and landmark point detector using MobileNet architectures and different loses for fast and accurate results on the mobile.
• Result: Successfully imitated the person on their avatar with high resemblance possible in a 2D environment with high aesthetic constraints.

May 2018 – Jul 2018 · 2 mos · Noida

• Work:-
• As a data scientist, I was supposed to build machine and deep learning models to help tone down, by identifying key factors, high level of attrition rates and build dynamic dashboards. Through this, the HR team monitored interpersonal relationships of employees and managers along with employee satisfaction rates, and, the Sales team can monitor their relationship with clients and revenue and counter any lapses in communication.
• Worked on sales pipeline visualization development that tracked various performance parameters of the marketing team as sales made, client feedback, etc. and developed expense visualizations. Also, devised a way to identify and fix communication lapses by implementing a case tracking system.
• Worked on a simple deep learning Chabot for company policy retrieval.
• Worked on development of an OCR tool that converts images to text through OCRSpace api.
• Result:-
• Reduction in the ongoing high level of attrition rates of the client, by using models like decision trees, random forests and neural networks, by 20% and better efficiency and communication among departments.
• OCR tool used to convert old printed documents to text.
• Recognition:-
• I was asked to manage interns that were hired subsequently.