AI Software Engineer
Amphenol · Pennsylvania
Industrial vision, manufacturing ML, and internal AI tools.+ AI SOFTWARE ENGINEER · PENNSYLVANIA, US
Intelligence.
Engineered.
I build systems that see, reason, and work.
From the first experiment to the last mile
of production.
OBSERVE · DECIDE · ACT
Computer vision, language systems,
and production engineering.
LEARNING SPACE / 01
POLICY NETWORKobservation → action
AGENT_01circuit simulation
ARCHITECTURE STUDIES / 3D
Drag to orbit · Scroll to zoom · Esc to exit
COMPUTER VISION+LLM SYSTEMS+SOFTWARE ENGINEERING+DEVOPS & MLOPS+SPATIAL COMPUTING
01 / THE ENGINEER
Curiosity in the lab.
Intent in the build.
I’m Shahab, an AI software engineer based in Pennsylvania. My work connects computer vision, language systems, and software delivery. I build research prototypes and tools used in production.
My Unity and XR background still shapes how I build: understand the environment, make interactions feel natural, and keep the system responsive.
Read my full resumeAI Engineer Intern
Intact Financial · Hong Kong
Document intelligence, geospatial ML, and CI/CD.ML & Computer Vision Intern
Molekiu · Hong Kong
OCR, GPU inference, edge models, and multilingual LLMs.Research Engineer · AR/VR & CV
The University of Hong Kong
Unity applications, spatial mapping, and a five-person research team.Software Engineer Intern
Hyr · Hong Kong
REST APIs, authentication, and backend performance.FOUNDATION
Applied Artificial Intelligence
The University of Hong Kong · 2022–2026
02 / SELECTED WORK
Inside the systems.
Explore ASCOTA, my computer vision pipeline for archaeological images, plus bank churn modeling and production AI systems.
EXPERIENCE / SYSTEMS IN PRACTICE
Inspect the architecture behind the work03 / THE LEARNING CIRCUIT
Change the objective.
Change the behavior.
A small reinforcement learning experiment.
Train an agent in your browser and see
what it learns to value.
REWARD OBJECTIVE
Make every step count.
Reach all four checkpoints with a penalty for each additional step.
AlgorithmQ-learning
Training episodes0 / 4,000
Evaluation steps—
Direction changes—
Checkpoints reached—
Ready. Choose an objective and run a training session.
REAL TRAINING · RUNS LOCALLY · NO SIGN-IN
Tabular Q-learning in a simplified grid circuit. The neural network in the opening scene is a visual motif, not this agent’s model.
04 / BEYOND THE MODEL
The last mile matters.
A useful model needs a dependable system.
I work across the software and infrastructure
that make AI usable.
01 Build02 Package03 Ship04 Improve
Software engineering
APIs, data models, and application logic that connect AI to the people using it.
- Node.js / Express REST APIs
- Python services & PostgreSQL
- Authentication & integration
HYR 20% higher backend throughput
DevOps & delivery
Repeatable deployment, containerized tools, and cloud workflows that move work beyond a notebook.
- Docker containerization
- GitLab CI/CD pipelines
- AWS / Azure deployments
INTACT 40% less manual deployment
Production ML systems
Model execution tuned for the constraints of real devices, factory workflows, and latency.
- ONNX & GPU inference
- Quantization & CUDA workflows
- Snowflake-backed ML pipelines
MOLEKIU Over 25% lower ML latency
ENGINEERING NOTES What I’m interested in building next +
Observable AI services
Trace a request from API to retrieval to inference. Track latency, cost, and failure modes together.
Evaluation in the delivery pipeline
Treat model quality as a release criterion, with regression datasets and repeatable checks before deployment.
Reliable agent backends
Typed tools, bounded retries, durable jobs, and human review at consequential decision points.
Future engineering directions, alongside the shipped work above.