Curriculum Vitae

Participated in developing multi-agent and task-specific skills for a privately deployed agent platform; packaged skills as standardized tools and configurations. Under NDA, business details are omitted.

• Supported packaging the project into OpenClaw for direct execution
• Designed input/output constraints and tool-calling boundaries to reduce format drift and invalid calls in multi-turn interactions
• Built a reusable tool adapter layer (APIs/files/data queries) with authentication, timeout control, retries, and exception fallbacks

Technologies: Python, Agent Engineering, Tool Calling, Workflow Orchestration, Observability

Date: Feb 2026

Participated in building an internal contract review assistant module that converts contracts into structured clauses and produces risk notes with evidence references. Under NDA, organization, contract types, and data details are omitted.

• Implemented contract parsing and clause indexing to produce structured outputs (clauses, original text locations, citation anchors) for traceable review
• Integrated retrieval and tool calling to bind review conclusions with evidence spans (paragraph/page), generating a verifiable risk list
• Packaged review task APIs and background processing (queue/retries) to support batch processing and resubmission after manual verification

Technologies: LLM Application Development, Document Parsing, RAG, Tool Calling, Multi-Agent Workflows

Date: Jan 2026

Participated in exploring 3D-enhanced scientific image forgery detection by introducing pseudo-3D and physics-inspired features into pixel-level segmentation, with supporting feature caching and evaluation workflows. Under NDA, data and business details are omitted.

• Implemented offline extraction and cache reuse for pseudo-3D and physics-inspired features to reduce repeated computation during training
• Improved data preparation and mask alignment scripts to support batch preprocessing, patch-based training, and unified evaluation entrypoints
• Generated inference visualizations and offline evaluation reports to compare the impact of different feature combinations

Technologies: PyTorch, Swin Transformer (Training), OpenCV, Feature Caching, Segmentation Evaluation & Reporting

Date: Dec 2025

Participated in building a scientific image integrity screening module that detects suspected copy-move operations and outputs pixel-level localization masks to assist reviewers. Under NDA, project source and data details are omitted.

• Implemented high-resolution sliding-window inference and mask stitching to output overlays, masks, and key region coordinates
• Implemented post-processing and risk grading strategies (thresholding, area ratio, etc.) to enable fast triage and ranking
• Provided FastAPI inference endpoints and batch scripts for offline runs, result persistence, and report generation

Technologies: PyTorch, timm, OpenCV, FastAPI, ONNX/ONNXRuntime (Deployment), Sliding-Window Aggregation, Image Forensics

Date: Nov 2025

Participated in developing a platform to assist review of scientific PDFs and images, covering document parsing, text-image extraction, similarity screening, quality evaluation, and report export. Under NDA, clients and data sources are omitted.

• Implemented an upload and parsing pipeline: normalized document formats, performed page-by-page parsing, extracted images/text, and generated traceable metadata (unique ID, page number, context text, etc.)
• Implemented “evidence filtering → parallel analysis”: filtered non-evidence images via a hybrid vision/rule strategy and ran quality evaluation, image-text consistency checks, and anomaly trace detection in parallel
• Implemented duplication and tampering risk detection: combined retrieval-based similarity, local matching, and deep model inference to output verifiable localization evidence and structured summaries (for frontend display/report export)

Technologies: Python, Document Parsing, Image Processing, Vector Retrieval/Similarity Search, Deep Learning Inference, Async Task Orchestration, Docker

Date: Oct 2025

Participated in developing a document processing and review platform covering upload, preview, OCR, information extraction, and export. Under NDA, organization, data, and details are omitted.

• Implemented end-to-end task orchestration and state tracking for long-running document jobs, including progress reporting, retries, step-level reruns, and checkpoint resume
• Implemented pluggable integration for document preview and parsing services via a unified adapter layer (request/response schemas, error codes, and timeout policies), enabling parser switching by scenario
• Built a multi-role Agent/workflow execution framework based on LangChain, persisting key intermediate artifacts in structured form and recording auditable logs for review, tracing, and reproduction

Technologies: FastAPI, React/Vite, Docker Compose, LangChain, OCR, Multi-Agent Workflows, RAG, Pluggable Toolchain

Date: Sep 2025

Extension of smoking detection research focusing on improved object detection in challenging lighting conditions. Developing a YOLO-based architecture with specialized components addressing the unique challenges of low-light environments. GitHub

• Designing custom loss functions specifically optimized for low-light object detection scenarios
• Implementing attention mechanisms to focus on key visual features in varying illumination conditions
• Optimizing upsampling techniques to preserve fine details in dark environments
• Integrating lightweight low-light enhancement module into the detection pipeline

Technologies: PyTorch, YOLO, Computer Vision, CUDA, Attention Mechanisms

Status: Ongoing (April 2025)

Related publications: YOLOv11‑LCDFS: Enhanced Smoking Detection With Low‑light Enhancement (manuscript in preparation)

Developing medical image analysis systems for cancer research. GitHub

• Designing medical image segmentation algorithms for hepatocellular carcinoma and renal cell carcinoma
• Designing methods for integrating clinical tabular data with imaging features for comprehensive analysis
• Implementing multi-modal fusion techniques for combining different CT scan phases
• Developing 3D volumetric segmentation approaches for comprehensive anatomical analysis

Technologies: Python, Deep Learning, 3D Segmentation, Multi-modal Fusion, PyDicom, NumPy

Status: Ongoing (April 2025)

Related publications: Multi‑modal medical image segmentation and fusion (manuscript in preparation)

Developed deep learning models for automated disease detection and classification from medical imaging data, with a focus on diabetic foot ulcer detection. GitHub

• Implemented and compared multiple YOLOv8 architecture variants enhanced with advanced attention mechanisms (GAM, CBAM, ECA, CoordAtt) for precise localization of diabetic wounds in clinical images
• Designed and evaluated custom YOLOv8 architectures integrating novel upsampling techniques including CARAFE and DySample for improved feature map resolution
• Developed innovative triplet-based loss functions and Inner-CIoU mechanisms to enhance detection accuracy for wounds of varying sizes and appearances
• Employed dynamic convolution techniques to adaptively capture wound features across diverse clinical settings and lighting conditions
• Constructed and curated specialized datasets of diabetic foot ulcers for training and comprehensive validation
• Achieved significant improvements in both detection accuracy and inference speed compared to baseline models, with particular gains for small and atypical lesions
• Implemented model explainability techniques to visualize feature importance and attention maps for clinical interpretation

Technologies: PyTorch, YOLO, Attention Mechanisms (GAM/CBAM/ECA), Custom Loss Functions, Feature Upsampling Techniques, Dynamic Convolutions

Date: August 2024

Developed an advanced machine learning system for Twitter content quality assessment and spam detection, achieving over 88% accuracy in distinguishing between legitimate tweets and spam content. GitHub

• Implemented comprehensive data preprocessing techniques for Twitter data, including text normalization, feature extraction, and handling of missing values
• Engineered complex features by combining user metrics (follower count, following count) and behavioral patterns to enhance classification performance
• Applied sentiment analysis and content analysis techniques to identify quality patterns in tweets
• Developed and compared multiple machine learning models to optimize classification accuracy
• Created data visualizations to communicate findings and identify key patterns in Twitter content quality assessment

Technologies: Python, NLTK, scikit-learn, Pandas, Matplotlib, Seaborn

Date: August 2024

Developed a comprehensive data analysis and machine learning system to predict Airbnb listing prices in New York City based on various property features and location data. GitHub

• Performed extensive exploratory data analysis on NYC Airbnb dataset with over 48,000 listings, creating visualizations to reveal pricing patterns across neighborhoods
• Implemented geospatial analysis to visualize property distribution and identify high-value areas using Python GIS libraries
• Engineered relevant features by transforming categorical variables and creating new metrics to better capture pricing factors
• Developed a RandomForest regression model to predict continuous listing prices with optimized hyperparameters
• Created a classification model to categorize listings into price brackets, achieving high accuracy through model tuning with RandomizedSearchCV
• Built interactive visualizations to help hosts and travelers understand pricing determinants in the NYC short-term rental market

Technologies: Python, Pandas, Scikit-learn, GeoPandas, Matplotlib, Seaborn, RandomForest

Date: January 2024

Learning to analyze airline customer sentiment from Twitter data using natural language processing techniques. GitHub

• Studying text preprocessing methods for social media data (handling hashtags, mentions, emojis, URLs)
• Learning about different feature extraction approaches including TF-IDF vectorization and N-gram analysis
• Working with various classification models and comparing their performance
• Practicing error analysis to understand patterns in misclassified tweets
• Creating visualizations using Matplotlib, Seaborn, and Plotly to understand sentiment trends

Technologies: Python, NLTK, spaCy, scikit-learn, Pandas, Matplotlib, Seaborn, Plotly, WordCloud

Date: December 2023

Undergraduate thesis project combining YOLO object detection with MediaPipe skeletal tracking to accurately identify and classify smoking gestures in video streams. GitHub

• Utilized YOLO object detection to identify cigarettes and related objects in video footage
• Implemented MediaPipe for real-time skeletal tracking and pose estimation
• Designed algorithms to recognize characteristic smoking hand-to-mouth gesture patterns

Technologies: YOLO, MediaPipe, Pose Estimation, Action Recognition, PyTorch, OpenCV

Date: April 2023

Outcomes: Patent application (202310277784.1), 3 software copyrights, thesis received outstanding evaluation

Learning to analyze health indicators for heart disease prediction using machine learning approaches. GitHub

• Working with health metrics from the CDC's BRFSS survey dataset
• Learning about feature selection and preprocessing for health indicators
• Studying both traditional machine learning models and deep learning approaches
• Practicing model evaluation using cross-validation and performance metrics
• Creating visualizations to understand relationships between health factors

Technologies: Python, TensorFlow 2.11.0, scikit-learn, Pandas, NumPy, Matplotlib, Seaborn

Date: May 2022