Projects
Software, computer vision, and algorithm projects, alongside selected professional work.
Project galleries
Explore the original collections, with project images, animations, and implementation details.
Deep Learning/CV
Learning based computer vision
3D Computer Vision
Learning for 3D Vision
Robotics & Autonomy
Perception, Planning, Controls
All projects
The complete repository collection, including the individual projects in the galleries above.
- A Star Path Planning
- BFS 8 Puzzle
- Camera Pose Estimation
- Control of Crane System
- Dijkstra Point Robot
- Generative Adversarial Networks
- Image Stitching
- Implicit Neural Representation
- K Means Clustering
- Lane Detection
- LiDAR SLAM
- Neural Radiance Fields
- Neural Surfaces
- Obstacle Avoidance Bot
- PointCloud Classification and Segmentation
- Point Cloud Analysis
- Pose Estimation 2D
- PyTorch3D Rendering Basics
- Semantic Segmentation and Object Detection
- Single View to 3D
- Stereo Depth Estimation
- Structure from Motion
- Superpixels and SLIC
- Trajectory Tracking
- Tsai Camera Calibration
- Wall Following Robot
- Zhang Camera Calibration
- manipulator control DDPG
Professional work
Selected applications, developer tools, and data infrastructure from NBCUniversal and FOX Sports.
NBCUniversal
Full-stack estimating & takeoff
A Python and React/TypeScript platform that combines document ingestion, OCR, spatial registration, and AI-assisted extraction.
- Documents
- OCR / registration
- Extraction
- Geometry
Conceptual flow based on the project description.
- Independently designed and built the platform, including a geometry-aware computation engine.
- Cut takeoff time from 6 hours to 30 minutes, backed by approximately 900 automated tests.
NBCUniversal
Production CI debugging agent
A production developer tool using Python, FastAPI, GitHub Actions, and LLM tool calling to analyze failed builds, identify likely root causes, and recommend code fixes.
- Failed build
- Analysis
- Likely cause
- Suggested fix
Conceptual flow based on the project description.
- Built and deployed the agent to analyze failed builds and recommend code fixes.
- Combined build diagnostics, likely-root-cause analysis, and LLM tool calling in the debugging workflow.
FOX Sports
Video understanding & media pipelines
Production video understanding and orchestration pipelines generating 100K+ searchable clip embeddings monthly, contributing to 30% higher retrieval accuracy.
- Video
- FFmpeg
- Embeddings
- Retrieval
Conceptual flow based on the project description.
- Built video understanding pipelines using Python, FFmpeg, CLIP/SigLIP, and TimeSformer; integrated FAISS/HNSW and BigQuery for semantic retrieval.
- Deployed Apache Airflow DAGs on GCP for 100K+ video segments monthly, with preprocessing, transcription, embeddings, and BigQuery ETL.
- Parallel execution, automated retries, SLA monitoring, and idempotent backfills achieved 99%+ pipeline success and 50% fewer manual interventions.
FOX Sports
Speech transcription & diarization
Speech processing pipelines handling 800+ hours of audio monthly, contributing to a 40% reduction in manual media search effort.
- Audio
- Whisper / VAD
- Speaker labels
- Transcripts
Conceptual flow based on the project description.
- Engineered production transcription and speaker diarization using Whisper, VAD, and speaker-embedding clustering.
- Generated timestamp-aligned, speaker-labeled transcripts and multilingual metadata for the media search platform.
NBCUniversal
Legal retrieval & evaluation
Retrieval and evaluation infrastructure over 5,000+ pages of legal documents, with semantic search and automated citation validation.
- Combined semantic search, automated citation validation, and deterministic checks.
- Achieved full citation coverage at an approximately 2% review-error rate.
FOX Sports
Event-driven operational monitoring
An event-driven monitoring service that surfaces ML pipeline drift alerts, exceptions, and health insights in real time.
- Built the service using AWS Bedrock, Lambda, and the Slack API.
- Reduced issue-resolution time by 35%.
FOX Sports
LLM evaluation & serving
Evaluation and regression tooling that improves retrieval quality, alongside serving optimizations that reduce per-request cost.
- Built an automated harness using synthetic queries, oracle sets, and precision@k; improved top-5 retrieval by 12 percentage points.
- Used response caching and lightweight request routing on Databricks to reduce per-request serving cost by 28% without quality loss.