Harishma Dayanidhi

Harishma Dayanidhi

About

Detail

Co-Founder
San Francisco, California, United States

Timeline

work
Job
school
Education
folder
Project
flag
Award

Résumé

Jobs verified_user 0% verified
  • Voxel
    Co-Founder and VP of Engineering, AI
    Voxel
    Jun 2021 - Jul 2025 (4 years 2 months)
    - Architected an AI system for real-time detection of 25+ events in live video (fastest time to value in market) - Built and led a 15 person Perception team with top-tier talent from Google, Meta and CMU - Deployed at multiple Fortune 500 customers and secured $35M in venture funding - Achieved hypergrowth with strong margins and high net retention
  • Voxel
    Co-Founder and VP of Engineering Perception
    Voxel
    Jun 2021 - Jun 2025 (4 years 1 month)
    - Architected an AI system for real-time detection of 25+ events in live video (fastest time to value in market) - Built and led a 15 person Perception team with top-tier talent from Google, Meta and CMU - Deployed at multiple Fortune 500 customers and secured $35M in venture funding - Achieved hypergrowth with strong margins and high net retention
  • AURORA
    Senior Software Engineer
    AURORA
    Sep 2019 - Jun 2021 (1 year 10 months)
    - Developed geometric methods to automatically expand canonical 3D bounding boxes into their full observed LiDAR shapes, generating accurate 3D ground-truth labels for object detection models. - Led development of a framework to detect data gaps in perception models and automatically surface high-value samples, improving model robustness and reducing labeling load.
  • Uber
    Software Engineer 2
    Uber
    Feb 2017 - Sep 2019 (2 years 8 months)
    - Invented and patented a novel motion-planning architecture for AVs focused on discrete decision-making under uncertainty. - Trained and deployed an XGBoost-based behavior prediction model for discrete actor decisions, enabling more efficient and interpretable motion planning. - Built an interactive visualization tool to explain XGBoost inference outputs and feature contributions, improving model interpretability for planning engineers. - Designed and implemented a 3D–2D association pipeline using calibrated projections and Hungarian matching to generate consistent cross-sensor ground-truth labels for LiDAR–camera perception models. - Developed an experimental framework to generate 2D bounding-box labels from 3D annotations using a ResNet-
  • I
    NLP Engineer (Internship)
    InRhythm Inc
    May 2016 - Aug 2016 (4 months)
    • Built a system in Python to automatically attribute a news article to a public company using Stanford’s co-reference resolution module • Implemented a heuristic based module to extract the public companies relevant to a news article • Wrote a Python module to interface with the Stanford POS tagger that improved the performance of an existing sentiment analysis module by a factor of ten
  • Carnegie Mellon University
    Research Assistant
    Carnegie Mellon University
    Sep 2015 - May 2016 (9 months)
    As part of my research assistantship I work with the NSF Frontier project, which builds on recent advances in natural language processing, privacy preference modelling and crowdsourcing to Semi-automatically extract key privacy policy features from natural language website privacy policies, and present these features to users in an easy-to-digest format that enables them to make more informed privacy decisions as they interact with different websites.
  • Adobe
    Member of technical Staff 2
    Adobe
    Jun 2013 - Aug 2015 (2 years 3 months)
    Development of Adobe Digital Editions and Reader Mobile SDK(RMSDK). Adobe Digital Editions is a ebook reading tool currently available on Windows and Mac platforms. RMSDK is a powerful rendering engine capable of rendering PDF, Epub2 and a new standard based on HTML5 - Epub3.
Education verified_user 0% verified
  • Carnegie Mellon University
    Master’s Degree, Computer Science/Information Technology
    Carnegie Mellon University
    Jan 2015 - Dec 2016 (2 years)
  • National Institute of Technology Tiruchirappalli
    Bachelor's Degree, Computer Science
    National Institute of Technology Tiruchirappalli
    Jan 2009 - Dec 2013 (5 years)
    1. Member of Spider club - technical club that conducts various technical workshops and undertakes projects and research in the field of electronics and various disciplines in computer science. 2. Member of Leo Club - involved in organizing various social events in the campus 3. Member of Athletics team 4. Event manager of 'Fix the android' - an event to identify and fix bugs introduced in the circuit of robots built using logic gates, to perform a certain pre determined task. The event is conducted as part of Pragyan ( An international Techno-management fest by NIT Trichy) 5. Head of guest lectures at Vortex 2013
  • Carnegie Mellon University
    Carnegie Mellon University
    Carnegie Mellon University
  • National Institute of Technology Tiruchirappalli NIT Trichy
    National Institute of Technology Tiruchirappalli (NIT Trichy)
    National Institute of Technology Tiruchirappalli NIT Trichy
Projects (professional or personal) verified_user 0% verified
  • B
    Behavioral Cloning
    Apr 2017 - May 2017 (2 months)
    Build, a convolution neural network in Keras that predicts steering angles from images
  • B
    Building predictive client-side profiles for personalized advertising
    Feb 2016 - Apr 2016 (3 months)
    Built a predictive online advertising system based on a machine learning model that does not store user data on the server side
  • D
    Dynamic Memory Allocator
    Nov 2015
    A general purpose memory allocator was implemented in C with capabilities to handle malloc, calloc, realloc and free requests. Optimized for space and efficiency constraints using segregated free lists and boundary tag coalescing.
  • T
    Tiny UNIX Shell
    Oct 2015
    A shell was implemented in C that emulated the functionalities of a UNIX shell such as support running programs in foreground and background mode and switch between them seamlessly, handling SIGCHLD, SIGINT signals with custom signal handlers.
  • D
    Determine correlation of data based on Inversion Attack on Machine Learning Algorithms
    Sep 2015 - Dec 2015 (4 months)
    Using Inversion Attack as the theoretical foundation, experiments were run on various Machine learning algorithms implemented using Orange APIs, to identify attributes that are closely related especially attributes closely related to the sensitive attributes.
  • D
    Dynamically notifying users of sensors in the IOT Context
    Sep 2015 - Dec 2015 (4 months)
    With the advent of Internet of Things, it is not an overstatement to say every device can potentially collect user data which might not always to desirable. To address this issue we built a mobile app that dynamically notifies the users of any kind of data collection around them, followed by a short summary of the data being actually collected.The app is currently supported on iOS platform.
  • F
    Fog Computing
    Jan 2013
    Implemented the mechanism to Detect and Prevent Insider Threat Attacks in Cloud.
  • A
    A Hybrid protocol to Secure the Cloud Against Insider Threats
    In this project we proposed and implemented a Hybrid protocol that uses Selective Encryption with data cleaning, Enhanced Neural Network based user profiling and decoy technology to combat the insider threat attack in the cloud. This work was published as a poster in IEEE CCEM (Cloud Computing For Emerging Markets) 2014.
Awards verified_user 0% verified
  • G
    Panelist at GHC 2024
    Grace Hopper Celebration
    Oct 2024
  • Forbes
    Featured in Forbes as co-founder of Voxel
    Forbes
    May 2022
  • Carnegie Mellon University
    Tartans on the Rise 2024
    Carnegie Mellon University
    Tartans on the Rise 2024
Publications verified_user 0% verified
  • A
    A Hybrid protocol to secure the cloud against insider threats
    A back-propagation neural network was trained to detect malicious user activity in a cloud environment. A user was to be identified using pre-determined factors such as commands used, amount of data download and type of data downloaded. User data consisting of five unique users was logged for a period of two days and finally used to generate a 10 dimensional input vector. The neural network used pattern recognition algorithms to indicate if it can deterministically determine any of the five users. The neural network was simulated in octave
This is a community-created genome.