Mrityunjay Kumar Email :
¯/mj2030 </mrityunjaykumar911 Mobile : +1-631-710-1058
Stony Brook University New York, USA
Degree: Master of Science in Computer Science Jan-2019 May-2020
Maulana Azad National Institute Of Technology Bhopal, India
Degree: Bachelor in Technology with major in Computer Science & Engineering July-2010 April-2014
Programming Skills
Programming Languages: C++, C, Java, Golang, JavaScript, SQL
Technologies: Apache Storm, Apache Spark, Redis, REST,Apache Kafka, AWS, Git/Perforce, MATLAB
Relevant Coursework: Distributed Systems, Operating Systems, Analysis of algorithms, Visualization
Microsoft May’2022 till
Software Engineer-2 Mountain View, CA
Snap Service
Implemented new lightweight service layer to provide native renderer workers to download from Storage service
to Azure blob cache to improve streaming latency by 25%.
Implemented testing framework to improve service reliability score by 85% and improved deployment
frequency by 10%.
VMware Inc. July’2020 May’2022
Member of Technical Staff-III Palo Alto, CA
Virtual distributed file system, Storage layer
Conceptualized and implemented distributed FSCK tool for analyzing and repairing storage metadata. This
tool is leveraged to analyze the state of metadata storage on disk after crash recovery by forming analysis
matrices and cross consistency patterns between multiple metadata data structures like B+ trees, bitmap,
segment usage table
End-to-end implementation of a cloud-native microservice to validate the integrity of the file system k-v store
(such as B+ trees). The service can be scheduled or run on-demand to check the health of the storage system
k-v store. In addition, the service is also used for other storage features, such as snapshots, un-map, segment
cleaning to verify consistency and integrity of the metadata.
Implemented aggregated snapshot capacity from scratch to support statistics collection with p99 latency in
range of 30 ms.
Distributed Systems Lab - Stony Book University June’2019 May’2020
Graduate Research Assistant, Prof. Shuai Mu Stony Brook, NY
Distributed Multi-core Transactional Database Engine
Implemented asynchronous replication with multi-process paxos to achieve minimal loss in throughput
Designed multi-core log truncation mechanism to support check-pointing for the transaction recovery protocol,
which allowed read only transactions to scale 10M ops/sec.
Implemented replay protocol to guarantee serializibility and consistency check pipeline for in-memory
streams & disk logs.
Implemented verification pipeline for generating CPU/Heap/Memory throttling graph using gperftools &
mutrace tool.
Implemented fast header-only/compiled, C++ logging library facilitating aligned memory allocators with
features like Rotating log and auto-flush to make transparent transaction serialization
Talentica Software April’2016 Jan’2019
Senior Software Engineer Pune, India
Estimation & Prediction algorithms for Wireless Systems
Designed & Implemented a stream data pipeline for machine learning inferences & predictions using Apache
Storm, Apache Kafka, AWS, Cassandra.
Improved network traffic prediction accuracy to 80% using important native SLA metrics, RF Frequencies &
feedback loop from auxiliary methods.
Implemented feedback pipeline to improve accuracy of indoor location prediction algorithm by 95% for live
BLE assets.
Developed a multi-modal Machine Learning model for network traffic classification for Audio and Video
streaming using Ensemble of regression and auto-encoders.
Added an ad-hoc client to Storm topology to support collection of Real-time data to Disk & Batch pipeline
using Java, Apache Spark.
Developed a live BLE asset view portal to support data collection team, adding 70% more correctness to site
calibration using JavaScript, Apache Kafka, Python.
Machine Learning Algorithm object storage service
Designed continuous delivery pipeline for Machine Learning models using gRPC, protobuf, Redis, RabbitMQ,
AWS S3 improving deployment frequency by 40%.
Wrote polyglot client suite in Java, python, Go, C++ to facilitate generic object serialization which lead to
improved continuous deployment frequency.
File Sync Application
Implemented delta file sync Service in C++ using native git-tree diff algorithm which reduce load in sync server
by 400%.
Developed an update framework to support release notification & binary upgrade for multi-talent architecture
using Java, C++, Python.
Designed rate limiter service to prevent throttling of indexing layer & sync layer, improving average P99 sync
latency by 30%.
Financial Document Search Engine:
Improved search relevancy by 45%. Improved Online Ontology enhancer by 30% via optimizing Neo4j query
consumption, NLP Pipeline for Document Clustering, Keyword Extraction, Text Classification.
MediaTek Aug’2014 April’2016
Software Engineer Noida, India
As a feature owner of Audio module, implemented design and bug-fixes for new product release. Improved
playlist design to reduce playlist update overload by 25%, Integrated BT Stack in MMI Layer
Wrote Software Layer code for Stress Test based Combo GPS-WiFi-BT Tool in VC++
Raft - A distributed fault tolerant in-memory key value database
- Implemented Sharded and replicated fault-tolerance key-value store based on raft protocol from scratch in Go.
- Tweaked leader’s election to ensure progress in case of repetitive failure in re-election step
Map-Reduce library Implemented distributed map reduce library and worker failures from paper.
Backup File System in Linux Kernel Implemented stackable virtual file system with Custom Visibility,
Retention and Version Management Policy to support backup to flat files under guidance of Prof. Erez Zadok using C
and Linux VFS Layer.
Encryption based System tool for Linux Kernel Implemented system call for encrypt/decrypt file using AES
algorithm provided by Linux kernel crypto API under guidance of Prof. Erez Zadok in C.
Learning to Fingerprint the Latent Structure in Question Articulation: In this paper, we show that the latent
patterns of questions can be represented as a system that maximizes a cost function related to the underlying objective
and can be approximated to building a memory of patterns represented as a trained neural auto-encoder. [publication]