$ cat ./work/sequence.md # case study
Sequence
Applied AI Engineering & Workflow Automation
Sequence is the AI workflow and automation platform behind Zorel.ai's enterprise security products. It connects machine learning models, computer vision systems, backend services, and cloud infrastructure into production-ready enterprise applications — and it exists to answer one question: what does it take to move AI beyond research and into scalable, real-world workflows? In production, that answer looks like 87% faster breach response across finance and enterprise networks.
Computer Vision
The heart of Sequence: real-time vision systems built for security, automation, and intelligent event detection — from raw camera frames to classified, scored, actionable events.
Threat Detection
Real-time pipelines for intelligent monitoring — recognizing human activity, scoring events, and classifying security incidents as they happen, with feedback-loop learning so the system improves on its own misses.
- —Human activity recognition
- —Behavioral & suspicious-event detection
- —Security event classification
- —Real-time inference & event scoring
- —Feedback-loop learning
Pose Analysis
Pose-based understanding systems that turn raw video into structured human movement: keypoint extraction, skeleton normalization, and temporal sequence models that classify what a body is actually doing.
- —Human pose estimation & keypoint extraction
- —Skeleton normalization & pose embeddings
- —Temporal transformers (MMAction2)
- —Ragged vs. padded tensor strategies
- —Multi-frame reasoning & motion understanding
Dataset Engineering
Models are only as good as the data underneath them. Sequence involved evaluating and rebuilding activity datasets — SPHAR, LIRIS Human Activities — with annotation strategies, versioning, and validation pipelines.
- —SPHAR & LIRIS dataset evaluation
- —Missing-data handling & preprocessing
- —Annotation strategy & dataset versioning
- —Feedback-loop pipelines & model validation
AI Research & Development
Deep learning approaches for vision-based understanding — carried from paper to production.
- Sequence & temporal transformers
- 3D action classification
- Temporal attention
- Representation learning
- Feature extraction
- PyTorch & TensorFlow
- MMAction2
- Custom architectures
- Training pipelines
- Hyperparameter tuning
Workflow Orchestration
The core thesis of Sequence: individual models are not products. Connecting them into intelligent business workflows is.
- Multi-stage AI pipelines
- Computer vision workflows
- Decision pipelines
- Real-time processing
- Event-driven automation
- API orchestration
- AI-powered business automation
Backend Engineering
Backend systems designed to serve AI reliably — not notebooks wearing an API costume.
- FastAPI & REST API architecture
- Authentication
- Service orchestration
- Business logic & workflow execution
- Event processing
Cloud Infrastructure
Cloud-native infrastructure built for machine learning workloads, from container to production inference.
- AWS deployments
- Docker
- Infrastructure automation
- Model serving
- Scalable inference
- Production deployments
Technologies
- PyTorch
- TensorFlow
- Transformers
- MMAction2
- Computer Vision
- Human Pose Estimation
- Action Recognition
- Deep Learning
- Python
- FastAPI
- REST APIs
- AWS
- Docker
- Cloud Infrastructure
- Model Serving
- Dataset Engineering
- Data Modeling
- Data Validation
- Training Pipelines
- Feedback Loops
Engineering Philosophy
“Rather than isolated models, Sequence demonstrates how intelligent systems operate reliably within enterprise software.”
Machine learning, computer vision, backend engineering, cloud infrastructure, and workflow automation — combined into complete, end-to-end AI systems. That is the standard I hold every engagement to.
Discuss a system like this