Selected Systems
From perception to production.
I build production vision systems and inspectable agent workflow references from perception and reasoning through runtime integration, evaluation, and handoff.
The projects below are grouped by evidence, not chronology. Together they show one operating pattern: make model behavior explicit, integrate it into a real system, measure failure modes, and leave a maintainable delivery path.
One delivery spine
- Perception: segmentation, matting, depth, detection, tracking, and visual understanding.
- Reasoning: composition state machines, agent workflow state, retrieval, structured decisions, and review.
- Runtime: quantization, TFLite/ONNX/TensorRT, C++/Android, GPU rendering, APIs, and services.
- Evaluation: hard cases, golden cases, visual comparison, traces, A/B tests, regression, and release gates.
- Handoff: technical scoping, debugging, documentation, reusable workflows, and production ownership.
Production foundation: Vision and Edge AI
A repeatable model-to-device path for camera and imaging systems.
System owned
Connected data preparation, training, quantization, export validation, native integration, device testing, and model replacement checks across segmentation, image matting, monocular depth, detection, and camera-oriented image processing.
Engineering judgment
Model architecture, I/O contracts, quantization behavior, runtime assumptions, and device debugging were treated as one design space rather than separate team boundaries.
Representative stack
Python, PyTorch, TFLite, ONNX, TensorRT/ONNXRuntime, OpenCV, C++, Android NDK, CMake, Docker, and device-side debugging.
Differentiation: Computational Photography
Perception, rendering, and camera behavior combined into controllable visual systems.
System owned
Worked across depth-aware bokeh, portrait segmentation and matting, subject protection, optical highlights, GPU rendering, and real-time composition using detection, optical-flow tracking, camera guidance, and application state behavior.
Engineering judgment
Final image quality depends on the interaction between model output, rendering parameters, long-tail boundaries, motion, device constraints, and review criteria. The work therefore included visual debugging, hard-case analysis, batch comparison, and release-quality review artifacts.
Representative stack
Monocular depth, image matting, detection and tracking, OpenGL/GLSL, C++ rendering, alpha compositing, OpenCV, TensorRT, ONNXRuntime, and Android camera workflows.
Adjacent growth: Reliable Agent and Multimodal Workflows
Tested orchestration mechanics and deterministic multimodal tooling, with explicit evidence boundaries.
Public evidence
Implemented a provider-neutral Python workflow for retry, fallback, structured-output validation, confidence gates, deterministic traces, and hard-case evaluation. Built a browser-local Photography Mentor for image diagnostics, retrieval, practice planning, and review state.
Engineering judgment
Routing and generation are not enough. Expected failures need explicit policy; invalid outputs must be rejected; low-confidence results require review; and every decision path should leave a trace that can be reproduced in a hard-case suite.
Representative stack
Python, dataclasses, provider protocols, validated output contracts, unittest, local retrieval, deterministic image analysis, and browser-local persistence.
Product system: Business Learning Studio
A privacy-first, offline-capable workspace that turns ambiguous business questions into evidence, experiments, and decisions.
Public evidence
Built a complete browser-local learning loop with diagnostic planning, four practice modes, transparent feedback, mastery tracking, 32 original knowledge cards, and a five-gate decision coach.
Engineering judgment
The product treats learning as observable performance rather than content completion. User state remains in the browser, runtime dependencies are vendored, and release automation rejects screenshots, local paths, private-source markers, and non-allowlisted files.
Representative stack
Progressive Web App, service-worker offline cache, local storage, deterministic coaching rules, generated JSON knowledge artifacts, Node.js contract tests, privacy audit, and GitHub Actions.
Operating method: Evaluation and Handoff
The common method that connects vision, runtime, and agents.
Translate ambiguous product or field issues into a technical scope, failure taxonomy, hard-case set, prototype, measurable acceptance criteria, regression path, and release decision. Preserve reproducible traces and turn working patterns into reusable tools, documentation, and handoff material.