Portfolio of
ISHWAR SONI
Computer Vision & Motion Processing Engineer
Specializing in human motion synthesis and processing large-scale datasets (AMASS). I build robust preprocessing pipelines and intelligent systems that bridge research and production.
Engineering Mindset
I enjoy solving confusing technical problems more than building demo projects. My work revolves around making messy systems reliable, whether it's normalizing AMASS motion datasets or deploying prediction models to production.
/Engineering mindset: I prioritize reliability and reproducibility.
/Debugging complex systems: I dive deep into coordinate transforms and data inconsistencies.
/Working with messy data: I build robust pipelines to handle real-world edge cases.
/Shipping over showing off: I value deployed, working systems over theoretical perfection.
Experience
Semantic Labs
Worked on large-scale AMASS motion datasets and BVH ↔ SMPL-H preprocessing pipelines.
- Built and optimized BVH ↔ SMPL-H preprocessing pipelines for large-scale datasets.
- Fixed critical coordinate system normalization issues affecting downstream models.
- Resolved orientation flips and scaling bugs in motion data.
- Implemented motion stabilization and smoothing algorithms.
- Optimized data pipeline performance for faster processing.
Featured Projects
A collection of technical case studies focusing on system architecture, data pipelines, and scalable infrastructure.
BVH-to-SMPLH Motion Processing Pipeline
Technologies
End-to-end pipeline to convert and clean motion capture data (BVH → SMPL-H) with correct orientation, grounding, and batch processing.
The Challenge
Raw BVH motion capture data suffers from orientation mismatches, floating/sliding feet, jitter artifacts, and lacks compatibility with SMPL-H based ML workflows.
The Solution
Implemented Savitzky-Golay smoothing for jitter reduction, FK-based grounding to fix floating characters, and foot-lock postprocessing to eliminate sliding artifacts. Built batch processing tools for large-scale datasets and reverse export (NPZ → BVH) for Blender/Unity-compatible workflows. Created MP4 preview generation and skeleton/frame inspector utilities for rapid validation.
Project Impact: High-quality motion data is the bottleneck for training human motion generation models. This pipeline automates weeks of manual cleanup and enables reliable, reproducible dataset preparation at scale.
End-to-end ML production system for real estate price prediction.
The Challenge
Existing models lacked a production-ready interface and deployment strategy.
The Solution
Optimized model serialization and implemented robust input validation in FastAPI.
Project Impact: Demonstrates full-stack ML engineering capability, from EDA to Dockerized deployment.
A reusable, modular machine learning pipeline architecture.
The Challenge
Spaghetti code in notebooks leading to data leakage and reproducibility issues.
The Solution
Implemented custom Scikit-learn transformers and pipelines.
Project Impact: Proves ability to write clean, maintainable, and leak-proof ML code.
High-volume corporate data extraction and structuring system.
The Challenge
Need to aggregate data for 10,000+ companies from diverse sources.
The Solution
Implemented retry logic, rotation, and robust parsing strategies.
Project Impact: Shows capability in data engineering and handling unstructured data at scale.
Technical Skills
Computer Vision & Motion
- Human Motion Analysis
- Pose Estimation
- SMPL
- SMPL-H
- BVH
Machine Learning & Data
- NumPy
- Pandas
- Scikit-learn
- Model Evaluation
- Data Engineering
Programming & Tools
- Python
- Git
- GitHub
- Jupyter Notebook
- Docker
- FastAPI