Masters in Computer Science
Graduation: May 2023
GPA: 3.8/4.0
New York University (Courant Institute of Mathematical Sciences)
View degree certificateAbout
Tejas Mahajan is a software engineer specializing in large-scale data systems and the infrastructure that runs them, with a focus on production-grade data platforms and real-time backends for compute-heavy, latency-sensitive domains. Currently at Nimble, I build the data and ML platform behind robot learning for an autonomous warehouse robotics company: streaming high-frequency sensor data off the robot fleet, turning it into reproducible training datasets for policy training, and developing the software that sets up and health-checks the stations that collect it. I also engineer the cost-aware pipelines and real-time services that turn fulfillment and fleet data into the operator-facing surfaces that run the warehouse floor. Previously at MerQube, I architected the equity and options data ingestion systems and unified data access layers powering complex index computation workflows. My background spans distributed systems, cloud-native infrastructure, and end-to-end ML deployment, from research-grade training and fine tuning deep learning models to deploying in resource constrained production systems. I'm particularly interested in building intelligent, reliable systems at the intersection of data infrastructure, robot learning, and autonomous decision-making.
Education
Bachelors in Computer Engineering
Graduation: May 2018
GPA: 3.8/4.0
University of Pune (Maharashtra Institute of Technology)
View degree certificateInterests
- Computer Vision
- Reinforcement Learning
- MLOps
- Agentic Systems
- Backend Software Development
- Generative AI
- Distributed Systems
- Large Language Models
Experience
March 2026 - Present
- Built a real-time data pipeline that streams high-frequency arm telemetry and multi-camera video from warehouse robot stations into a central store, so every demonstration can be searched within seconds of recording. Scaled it from one pilot station to the whole collection fleet, and the training corpus grew 4x within a month of rollout.
- Shaping the ML platform as a product for the robot learning team, turning researcher needs into the end-to-end workflow for filtering demonstrations, building versioned and reproducible training datasets, and tracking quality. After launch, training runs and trained models grew 5x within a month.
- Developed the hardware configuration and calibration modules for robot station setup and health checks. Stations now check their arms, grippers, cameras, pedals and motion trackers on their own, and every recording is tied to a verified calibration.
- Re-engineered a legacy BizOps warehouse-operations dashboard into a documented, stakeholder-aligned, and performant medallion pipelined datasets, codifying ad-hoc metric definitions into parity-validated bronze/silver/gold pipelines and cutting monthly compute spend roughly 10x.
July 2023 - March 2026
- Engineered the migration of equity reference and end-of-day pricing pipelines to a new provider platform, ensuring uninterrupted data delivery for index calculations with zero downtime.
- Led the development of a scalable options data platform, defining the data model, building ingestion and monitoring systems, and partnering with product, data providers, and financial engineers to resolve complex data integrity challenges. Delivered a unified data access layer that powers multi-asset index development and self-service analytics across teams
- Transformed operations workflow by developing automated validation checks for end-of-day prices and corporate action data across multiple providers. Enhanced data accuracy and consistency with customizable override capabilities.
June 2022 - Aug 2022
- Performed in-depth analysis of a vast flight data corpus to uncover spatial and temporal correlations, successfully identifying likely sources of false positives between airborne and ground-based objects.
- Fine tuned the intruder detection models by strategically integrating hard negative samples, leading to a decrease in false positives while preserving optimal sensitivity.
Data Scientist
Karza Technologies Pvt. Ltd. (Acquired by Perfios Technologies Pvt. Ltd.)Mumbai, India
Sept 2018 - Mar 2020
- Successfully delivered an end to end OCR Pipeline for the KYC Documents by implementing a robust synthetic data generation pipeline, model validation, hypothesis testing and training modules for text recognition, card detection and text detection tasks.
- Developed a light weight reformulation of fixed length text recognition model using the CTC framework by progressively optimizing the model towards bridging the latency vs task specific evaluation metric trade-off.
- Engineered the first deployment of deep learning models across AWS Lambda serverless functions and GPU instances, meticulously optimizing for compute, memory, and response time requirements
Aug 2018 - Mar 2019
- Implemented and fine-tuned the Retina-Net object detection network to detect 20 objects comprising garments and footwear. Obtained a mean average precision of 80 on the test dataset.
- Extracted the colour of detected object using image processing functions and mapped the colour to its appropriate shade name.
- Enhanced the user’s visual experience with detected objects image tags overlaid on the image and the generated metadata used to improve the underlying search algorithms.
Projects
Dec 1, 2025
Helios AI (xAI Hackathon)
Helios AI transforms solar farms from reactive assets into self-monitoring, self-diagnosing systems dramatically reducing inspection time and unlocking continuous, autonomous operations.
Dec 1, 2022
Exploring the Potential of Federated Learning for Medical Image Analysis in Non-IID Settings
Implemented a federated learning extension of ConVIRT to study privacy-preserving contrastive representation learning for medical image analysis under IID and Non-IID data distributions.
Dec 1, 2021
TEASEL: A Transformer-based Speech-Prefixed Language Model
Implementation of the TEASEL paper, demonstrating speech-conditioned Transformer prefixing for efficient multimodal sentiment prediction.
Implementation of a research paper.
Aug 1, 2020
Cartoonizer
An ML-powered app that stylizes real-world images and videos into expressive cartoon visuals.
It was used by over 200k+ users and to cartoonize a short film called The Black Disquistion.
May 1, 2018
Understanding Effective and Emotional Components in Advertisements
Explored neural architectures for predicting advertisement effectiveness and affective dimensions (arousal & valence) from multimodal visual data.
Accepted as a paper at the CVPR 2018 Ads Workshop.
Accomplishments
Deep Learning Auto Tagging Competition (HackerEarth)
Secured 1st place in a deep learning challenge to build a model that auto-tags gala images with relevant labels.
Udacity Artificial Intelligence Nanodegree
Udacity Deep Learning Nanodegree
Blog
Jun 17, 2021 · 1 min read
ExternalHow we built an inexpensive, scalable architecture to cartoonize the world!
Describes how we designed and deployed a low-cost, scalable system to serve ML-powered image and video cartoonization to users worldwide.
Jan 16, 2019 · 1 min read
MediumMachine Learning — Should you be a first mover or fast follower?
Explores whether in machine learning it’s better to innovate as a first mover or strategically enter later as a fast follower, weighing the advantages and drawbacks of each approach.
Contact
Email: tejas.mahajan121@gmail.com
Location: San Francisco, CA, USA