Linh
Tuan Ung

AI that scores how you say things, not what you say. In real-time.

• Built the WebSocket streaming path so pronunciation scores come back in under a second.
• Deployed and tuned edge ML models that classify phonemic variations.
• Secured and hosted the public API on a single edge GPU — that was the whole constraint.

Data moves between complex schemas. Nothing gets dropped. That's the job.

• Architected the core request routing for a multi-tenant publisher, maintaining zero dropped requests.
• Designed a schema-migration UI whose real job is stopping users from breaking their own database.
• Implemented strict RBAC, plus data validation that assumes every input is hostile.

Counting the pieces after a satellite breaks up. In simulation, for now.

• Generated a multi-camera synthetic dataset of a debris cloud in Blender, scripted end to end in Python.
• Found where YOLOv8 gives up on counting small, densely packed fragments: resolution limits and overlapping boxes.
• Swapped in CSRNet density estimation instead, then tested whether it held up on camera angles it had never seen.

Archive