Notes

Writing along the same system spine.

These notes develop reusable engineering judgment around perception, reasoning, runtime, evaluation, and handoff. They support the portfolio without repeating the CV.

Production foundation: Vision & Edge AI

Model quality and deployment quality are one design problem across data, architecture, export, runtime, and device constraints.

  • Debugging paths from PyTorch to TFLite and ONNX.
  • PTQ, QAT, operator compatibility, and INT8 regression analysis.
  • Validation checklists for native and on-device vision runtimes.

Differentiation: Computational Photography

The final image is shaped by perception, rendering, camera behavior, long-tail boundaries, motion, and review criteria.

  • Why AI bokeh is difficult at subject boundaries.
  • Engineering constraints of monocular depth in camera effects.
  • State-machine design for real-time composition guidance.

Growth vector: Agentic & Multimodal AI

The goal is not longer model output. It is visual and knowledge work that is constrainable, traceable, comparable, and improvable.

  • Designing local RAG and evidence-aware workflows.
  • Using schemas, routing, retries, and fallbacks for reliability.
  • Evaluating multimodal decisions with hard cases and traces.

Operating method: Evaluation & Handoff

This track covers the work between a promising prototype and a system another team can operate and improve.

  • Translating ambiguous feedback into hard cases and acceptance gates.
  • Building reproducible comparison reports and regression paths.
  • Codifying interfaces, assumptions, debugging paths, and release criteria.