New Delhi, Delhi, India
Building AI powered Backend Services at Microsoft, for providing customers quick and consistent experience for their Azure Compute tickets and live site incidents I have been Tech Lead for Large Scale projects impacting Millions if not Billions across the globe 1. Microsoft Copilot Mobile Vision : Android Tech Lead. DAU in Thousands, MAU in Millions 2. Motorola Deep Integration with MS Copilot : Worked on the Data Monitoring, Processing and Scheduling Pipeline for On-Device Recall (Retrieval Augmented Generation) 3. Microsoft Devices Ecosystem & Platform : Lead Test Engineer for the Security Key Provisioning and Attestation Framework delivered to customers like Jabra, MediaTek etc
Core Engineer for AI orchestration of quota and compute management operations at Azure Fleet (Azure Core). AI and Backend Engineering. Building a Copilot, that takes actions, saves time, interacts with stakeholders, and drives data driven decision making to solve Live Site Incidents and Quota Requests
Building Features for Microsoft Ai and Copilot App 1. Android lead for Mobile Copilot Vision. Shipped across the globe on Copilot app 2. Building Complex Search for LinkedIn. Web Lead and AI engineer 3. On-Device AI Models for Speech and LLM. Android Lead for Speech Pipeline - Diarisation, Translation, Transcription, Speaker Identification
📌 Core Engineer for designing and implementing audio pipeline for AI Companion App involving features such as real-time transcription, translation, diarisation, and speaker identification, incl. cloud and on-device frameworks. The latency achieved was 2.5 seconds 📌 Implemented LLM Orchestrator for AI Companion App to detect and take actions on user's intent, transcriptions and live screen recordings, wherein I performed backend development, prompt tuning and LLM chain creation. The Orchestrator response time was 2 seconds 📌 Core engineer for handling message (sms/mms/rms) data collection, scheduling (for data extraction, chunking, tokenizing) and processing (indexing to vectorDB) module to be used to deliver Message Recall functionality. Overall time of the pipeline was 7 seconds. Worked on delivering this functionality to Xiaomi and Motorola 📌 E2E Automation Architect and lead responsible for overlooking synthetic dataset creation, implementing UI automated validation framework for Microsoft Copilot recall scenario's and creating PowerBI dashboards 📌 Ai Accelerator: Prototyped image to GiF feature for Microsoft Expressive Input Panel
📌 Developed frontend for Bug Report Analyzer Web Application with functionalities such as Authentication, Forms and Dashboard Generation 📌 Worked on development of the Security Key Attestation and Provisioning feature in the Native(Java) and HAL(C++) Layer 📌 Responsible for performance analysis of the Display Subsystem (incl. Dual-display feature, SurfaceFlinger Composition and Rendering). Implemented a framework to automate this analysis for each release 📌 Implemented automated performance analysis tools to benchmark Software Health (performance, temperature, power and memory consumption) of the device. Incorporated them into a xTS tool which is shipped along with MS android platform. It has more than 200+ test-cases and is platform agnostic and feature dependent 📌 Prototyped hand gesture recognition Ai for Teams which performed computer based gesture detection using object detection algorithms in real-time (40ms)
📌 Modernized organization BOT application using latest dependencies pertaining to .Net Core & Microsoft BOT framework. Implemented new business logic which resulted in bug fixes, faster & better performance, and improved business logic handling. 📌 Implemented Orchestrator Language Recognition tool ( an offline NLP based model that uses KNN algorithm and provides faster training along with high prediction accuracy ) to enable horizontal scaling of user intent handling. 📌 Designed and Implemented business logic for Zero Code Onboarding & Refreshing/Updating of QnAMaker Knowledge Base. The manual process of onboarding knowledge base was reduced to an average of 90 seconds for 10K+ questions. 📌 Designed and Implemented a fully automated flow for Organizational product management tool access approval request management, which was previously manual and involved lots of back and forth mailing. 📌 Technologies Used : - C# Language - .Net Core 3 and Microsoft BOT Framework. - Azure Automated Flows : PowerApps, LogicApps, PowerAutomated - Database management and NoSQL : CosmosDB - API creation, management and security ( authentication & authorization ) - Microsoft Tools : SharePoint, Outlook, Teams, CPMT - Azure Administration - Azure Cognitive Services : Orchestrator, Dispatcher, LUIS, QnAMaker, TextAnalytics
Developed "Traffic Classification & Count Software" for Balaji Infratech. Problem Statement : To create an automated system to replace the current manual process of counting and classifying Indian Vehicular traffic over any sort of physical ( environmental - classification in both day and night conditions ) and technical constraints ( inappropriate lighting, not able to deploy state of the art camera's etc. ). Led a team of 30+ Software Engineering & Research interns to achieve the following development tasks : 📌 Data Set Creation : Created a annotated dataset of 20 Indian vehicular classes. 📌 Image Processing : Researched and implemented methods to improve Night time conditions for better object detection and classification. 📌 Object Detection & Classification : Researched and implemented various object detection models ( Yolo, Detectron, RetinaNet etc. ) to find the best model meeting our accuracy and processing time requirements. 📌 Object Tracking : Researches and Implemented various object tracking models ( such as DeepSort, Sort etc. ) to provide solutions to problems such as occlusions, fast moving vehicles etc. 📌 Frontend Development: Implemented v1.0 frontend using Python Tkinter framework. Mentored frontend development team to create best in class graphical user interface using Python - Kivy framework ( v2.0 ) The Software is able to achieve an accuracy of 70%+, with an average processing time equal to that of 50% of the Video Recording time ( time of Input Visual data for classification and counting ). The Software is successfully deployed at various locations of India. 📌 Technologies Used : - Python Language - Pytorch, Keras, Kivy, Tkinter & Tensorflow Frameworks.