Sep 2023 – Feb 2025 · 1 yr 5 mos · Milpitas, California, United States · Hybrid

Feb 2020 – Jul 2023 · 3 yrs 5 mos · Redwood City, California · On-site

• Early employee of a now unicorn startup that makes full stack software for warehouse robots. Delivered Segmentation and error detection systems, and built robot camera calibration tools.
• Conceptualized and deployed a new Error Detection System for parcel singulation robots at a top 3 logistics firm [Patented]
• The system, developed in Python, detected and flagged errors in real time (<100ms) on high speed conveyors to reduce parcel mis-slotting (from 5% to 0.01%) and avoid parcel-robot collisions
• Developed a semi-supervised learning approach in PyTorch; containarized the system and deployed it with NVIDIA Triton Server (using CUDA shared memory)
• Integrated the system with GCS to store live data and used Google BigQuery to generate comparison charts; created Kibana dashboards to visualize the live performance of the system.
• Created a 1-click self serve tool for robot camera calibration and verification using GRPC api in Python [Patent pending]
• Delivered automated processes for faster robot camera calibration (static as well as end-effector mounted cameras) and its verification getting sub-mm accuracy
• Reduced the calibration time by 10 mins and produced a comprehensive report
• Created a system diagnostic testing platform using GRPC APIs allowing engineering teams to create system tests at scale
• Initiated the effort; led meetings with Engineering, UI and DevOps to understand their pain points.
• Developed the backend using a graph based workflow which modularised the engineering, content and UX.
• Revamped segmentation pipeline for parcel singulation, helping Dexterity acquire 2 of its first 3 clients
• Revised data annotation by removing redundancy and used committee based active learning policy to select high-value samples and save cost
• Trained segmentation models in Pytorch to detect rigid as well as non-rigid parcels and used hard negative mining to improve MAP by appx. 15%; used GCS for AB testing of models and Plotly for visualizations.

Jul 2018 – Feb 2020 · 1 yr 7 mos · Sunnyvale, California, United States · On-site

• Amazon Halo (health tracker) - Worked on the Visual Body Fat (VBF) measurement feature
• Trained multi-view convolutional neural networks with supervised learning in PyTorch to predict body fat percentage from multi pose images; increased accuracy on the test set by appx. 25%; used AWS for multi-gpu training
• Identified issues with data annotation, worked with fitness experts and data team to make significant changes in the process. As a result, we hit our KPIs for bounding errors
• Amazon Ring cameras - trained neural networks in Keras for the human detection pipeline for indoor cameras
• Used model distillation and active learning to increase the mean precision over classes by 12%
• Worked with a team of engineers and QAs to iteratively improve the models and get it deployed in an Android app

May 2017 – Aug 2017 · 3 mos · Greater Boston

• I developed an augmented reality system which can project on real world environment in real time. The information regarding the environment comes through a camera. I developed and implemented a custom algorithm to calibrate the camera and projector.
• Key responsibilities:
• Took ownership of the project and developed it from scratch - from building the hardware setup to delivering a working demo.
• The projection was accurate and instantaneous for human perception.

Mar 2016 – Jul 2016 · 4 mos · Bengaluru, Karnataka, India

• I developed the Sync feature of Samsung SHealth android fitness application. This allows user to sync his/her fitness data on other android applications with SHealth.
• Exposure: Agile programming paradigm, Rest APIs, multi-threaded Java programming, object oriented design and architectures.

Jun 2014 – Mar 2016 · 1 yr 9 mos · Bengaluru, Karnataka, India

• Samsung Research Institute, Bangalore was my first company and it was a learning experience working with people here for two years.
• Developed the framework ‘S-Fit’ using R and Java to reduce user drop out in SHealth. The framework extracted useful patterns from the data coming from the wearable sensors and user food inputs using machine learning models to give recommendations. Ontology based knowledge was then used to cross validate the recommendations given.
• Published paper in IEEE, International Conference on Bioinformatics and Biomedicine [875-882].
• S-Fit: Knowledge guided fitness pattern mining framework, Ayush Sharma et al. DOI: 10.1109/BIBM.2015.7359800.
• Exposure: R, Data Analysis, Machine Learning, Java.

May 2013 – Jul 2013 · 2 mos · Bengaluru Area, India

• Worked on Content Based Image Retrieval(CBIR) Algorithms and how these can be adapted for android platform. Made an android application for the same.
• Exposure: CBIR, Java, Object Oriented Oriented Design, LIRE.