Hi everyone!

I am Siddhi and my passion for building machine-learning solutions has made me pick machine-learning engineering as a career. I have been working with machine learning, deep learning, computer vision, data infrastructure, and machine learning operations for over 6 years. I'm currently a Senior Machine Learning Engineer at ONRAMP, building recommendation and optimization systems for the freight industry. You can check out my projects here.

My first job was as a data scientist in one of the subsidiaries of a prestigious fintech company in Nepal called extensodata. At extensodata, I mostly tangled with huge structured fintech data ranging from banks to e-wallets. I began realizing that data science is a huge field and can get very vague.

I wanted to specialize in computer vision and joined Leapfrog, where I worked on some image segmentation and object tracking projects. After working for a year, I decided to continue my studies and joined the graduate program at the University of South Dakota.

• Languages: Python, Go, C, C++, SQL, Bash.
• ML / CV / Multimodal: PyTorch, PyTorch Lightning, TensorFlow, scikit-learn, Ultralytics (YOLO), OpenCV, TensorRT, depthai.
• Data Processing & Pipelines: Apache Airflow, Argo Workflows, PySpark, HDFS, Pandas, NumPy.
• Cloud & Infra: AWS (EC2, S3, EKS), GCP (GKE, Vertex AI), Kubernetes, Helm, Docker.
• MLOps & Observability: MLFlow, ClearML, Grafana, Sentry.
• Serving & APIs: FastAPI, Flask, Django Rest.
• Database: MySQL, MongoDB, SQLite, Postgres.
• Sensors & Edge Hardware: OAK Camera, NVIDIA Jetson, ESP32, Raspberry Pi.
• Version Control: Git, Bitbucket.

ONRAMP, New York, USA

• Designed the data architecture for a production recommendation system and built the Go and Python microservices serving recommendation and analytics workloads at scale.
• Migrated and consolidated ~10 legacy cloud functions into a generalized, parameterized Argo Workflows framework (self-hosted), eliminating VPC connector overhead and turning one-off featurization and data-preparation jobs into reusable, maintainable pipelines for downstream ML tasks.
• Containerized and deployed services with Docker, Kubernetes, and Helm on GCP, optimizing for scalability, compute utilization, and cost efficiency across the end-to-end pipeline.
• Built simulation workflows for rapid, pre-production experimentation, evaluating proposed recommendation and optimization algorithms against historic transaction data and projecting ~15% fuel cost savings for carriers.
• Provided technical leadership in the design and specification of optimization and recommendation algorithms, driving tradeoff analysis across quality, speed, and cost.

i8 Labs Inc, Mountain View, California, USA

• Trained, fine-tuned, and productionized custom YOLO computer vision models (PyTorch/Ultralytics, AWS) for real-world sensor streams, achieving an 8% increase in mean Average Precision (mAP) with metric-driven evaluation across model iterations.
• Built and standardized an internal semi-automated data annotation and curation pipeline to surface high-value training data, improving model accuracy and cutting manual labeling effort by 80%.
• Reduced Docker image size by 60%, cutting inference footprint, storage cost, and deployment time across a fleet of heterogeneous edge/IoT devices.
• Maintained 99% uptime for deployed inference services using Grafana for monitoring/observability and Balena for fleet device management, prioritizing reliability and operational simplicity.

MSc. Computer Science at University of South Dakota (Vermillion, USA)

• Took a break after a successful career in machine learning to pursue further academic studies.
• Published a paper at the IEEE AI Conference and co-authored a book on machine learning.

Leapfrog Technology, Kathmandu, Nepal

• Lead and mentored the AI/ML team for project delivery.
• Lead end-to-end client requirement elicitation process.
• Defined & developed standard ML practices.
• Used deep-learning frameworks such as darknet, OpenCV, and Tensorflow TRT to train & evaluate YOLO models for object detection.
• Build, Deploy and Maintain statistical, ML, and Deep learning using standard ML\MLOps frameworks such as MLFlow.
• Model tuning and optimization focused especially on deep learning models for embedded devices (NVIDIA Jetson Developer Toolkit).
• Worked on a multi-object tracking project using quantized YOLO tiny models for object detection, & deepsort for tracking.
• Worked on a prototype for a calorie estimator by segmenting the items on a plate using Masked RCNN.
• Experience working with human-computer interaction.
• Worked as team manager for the AI team.

Extensodata Pvt. Ltd, Kathmandu, Nepal

• Use and development of Data Architectures.
• Explanatory Data Analysis (EDA) in SQL as well as Jupyter notebooks.
• Using big data tools such as Hadoop, Spark, Hive, etc to manage huge volumes of data effectively.
• Data visualization using python libraries (seaborn, Matplotlib) and other third-party tools such as PowerBi & Apache Superset.
• Using various machine\deep learning models in spark (MLLib) as well as python (Sci-kit Learn, Keras).
• Using Pentaho and spark for extraction, transformation, and loading data from raw data (files, database, HDFS, hive) to required data architecture.
• Study feasibility, pros, and cons of machine learning and statistical models.
• Query optimization in a relational database (Mysql) for quicker data analysis.
• Writing automation scripts for various purposes (such as ETL, web scraping, etc) using python and Linux shell scripts.
• Studying the application of machine learning models in the banking domain.
• Generating and studying relevant using different feature engineering techniques (such as custom and quartile binnings, combining multiple features) in bank-specific data.
• Building prototype machine learning models on an ad-hoc basis as well as deployable backend data structures.
• Writing stored procedures and scripts to generate various reports from source data for UI consumption.
• Mentoring interns, trainees, and Junior members of the Team
• Part of the small team that built and launched Foneloan, Nepal's first collateral-free instant digital lending product, now offered by 11+ major commercial banks; built the data pipelines (Apache Airflow, Pentaho) automating loan-disbursement decisions with ML and statistical models, eliminating the human decision loop.
• Built an AWS-based OCR pipeline (Tesseract + Python) to automate text extraction from national identification cards, improving data entry speed by 90%.

Deep Spectral Features to Detect Atrial Fibrillation using Single-Lead ECG Signals
2023, 10.1109/CAI54212.2023.00074

Cracking the Machine Learning Code: Technicality or Innovation?
2024, 10.1007/978-981-97-2720-9

A Hybrid Transformer Model for Robust Multi-modal Emotion Recognition Using Audio and Text Data
2025, 10.1007/978-981-96-9533-1_22

MLOps | Machine Learning Operations by Duke University
Coursera (Aug 2024)
C2229E7K8YCJ

Container Orchestration using Kubernetes
Coursera (Aug 2024)
1df59f5037a370f3e6107779451aeca2

Introduction to Machine Learning in Production
Coursera (July 2022)
https://coursera.org/verify/AQYAFFTRKJW9

Optimize TensorFlow Models For Deployment with TensorRT
Coursera (June 2022)
https://coursera.org/verify/K343P63ZCMNR

Speak Like a Pro: Public Speaking for Professionals
Udemy (June 2022)
UC-9c99a21f-818b-43fa-9f06-34f457356a6d

Deep Learning Computer Vision™ CNN, OpenCV, YOLO, SSD & GANs
Udemy (May 2022)
UC-cea3a356-fa52-46a5-8120-f09bcea73506

Machine learning Deep Learning Model Deployment
Udemy (Oct 2021)
UC-cea3a356-fa52-46a5-8120-f09bcea73506

Applied Artificial Intelligence Club
President
SGA Club, University of South Dakota

University of South Dakota
MS in Computer Science (Aug 2022 - Dec 2023)
GPA: 4.0/4.0
Coursework: Computer vision, Machine learning and pattern recognition

Tribhuvan University
Bachelor in Computer Engineering, KEC (Tribhuvan University) (2014-2018)

See the full, categorized list in projects.md.

ML Research & Graph Learning

• Bluesky GNN — Graph neural network experiments on the Bluesky follow graph, crawled via the public AT Protocol API.

LLM, NLP & Multimodal

• RAG Document Q&A System — Retrieval-augmented generation over PDF/JSON/CSV/web documents using embedding-based vector search (ChromaDB), LangChain, and LangGraph.
• Multimodal Emotion Analysis — Multimodal network (LLMs + visual models) to identify emotion from speech and video streams.

Computer Vision & Edge AI

• OAK Dashcam — Multi-camera dashcam for Raspberry Pi + Luxonis OAK, running H.265 encoding and YOLO detection entirely on-camera so the Pi's CPU stays free.
• YOLO Ultralytics + ClearML — End-to-end YOLO training pipeline from CVAT annotations to a ClearML-tracked, S3-backed model registry.
• Multi Object Tracker — Multi-object tracking system on Jetson Nano using DeepSORT, YOLOv3, and OpenCV.
• Reidentification — Training and evaluation scripts for reidentification in multi-object tracking.

Recommendation Systems & Big Data

• Pyspark Recommendation System — Distributed book recommendation system using PySpark, MongoDB, and HDFS.

Healthcare & Signal Processing

• Arrhythmia Detection — Detection of arrhythmia in ECG signals using spectral representation and deep learning.

Real-World Products

• Foneloan — Nepal's first collateral-free instant digital lending product, now offered by 11+ major commercial banks.