1. Perception
Problems: segmentation and matting, monocular depth, object detection, optical flow, tracking, image-quality understanding.
Stack: PyTorch, OpenCV, image processing, data pipelines, training and validation tooling.
Technical Map
Perception to production.
Five connected stages describe the work more accurately than a flat skill list.
2. Reasoning
Problems: real-time composition state machines, provider-neutral agent workflows, local retrieval, structured decisions, failure policy, human review.
Stack: Python, dataclasses, provider protocols, validated output contracts, deterministic traces, local retrieval.
3. Runtime
Problems: quantization, export compatibility, native integration, rendering, device deployment, service packaging.
Stack: TFLite, ONNX, TensorRT/ONNXRuntime, C++, Android NDK, CMake, OpenGL/GLSL, Docker, FastAPI.
4. Evaluation
Problems: hard-case definition, model replacement, visual quality, provider comparison, agent stability, release decisions.
Stack: golden cases, batch simulation, visual montages, regression suites, traces, A/B reports, acceptance gates.
5. Handoff
Problems: ambiguous requirements, cross-team dependencies, field feedback, debugging, release readiness, maintainability.
Practice: technical scoping, prototypes, issue taxonomies, playbooks, documentation, reusable workflows, production ownership.
Positioning boundary
The production foundation is computer vision, computational photography, and edge deployment. Reliable agent and multimodal workflows are the adjacent growth direction. The public evidence supports orchestration, retrieval, validated outputs, failure handling, traces, and evaluation; it does not claim live commercial-provider integration, foundation-model training, or hyperscale LLM infrastructure.