About

From perception to production.

I build production computer-vision systems and public agent workflow references from perception and reasoning through runtime integration, evaluation, and handoff.

My production foundation is camera and edge AI. Computational photography is the differentiated domain; reliable agent and multimodal workflows are the adjacent growth direction.

Production foundation

Vision and edge AI: training, quantization, native runtime integration, on-device debugging, and release validation.

Differentiated domain

Computational photography: segmentation, matting, depth, tracking, rendering, and visual-quality evaluation.

Growth vector

Agentic and multimodal AI: explicit state, provider-neutral routing, validated outputs, failure handling, traces, and evaluation.

Selected Evidence

Two inspectable artifacts: one visual product and one tested reliability reference.

Photography Mentor public knowledge atlas with 62 source-linked cards

Public application

Photography Mentor

A commercial-ready, browser-local learning system combining 62 source-linked cards, deterministic image diagnostics, deliberate practice, review state, offline use, and portable learner data.

Image analysis62-card RAGOffline PWAData portability
Open applicationRead system notes
01 routing     primary    selected
02 generation  primary    started
03 generation  primary    exhausted
04 routing     fallback   selected
05 validation  fallback   passed
06 review_gate fallback   passed
07 completion  fallback   accepted

Public reference code

Reliable Agent Workflow

A provider-neutral Python reference for retries, fallback, validated output contracts, confidence gates, deterministic traces, and hard-case evaluation.

PythonTyped stateFallback10 tests
Review codeRead architecture

One System Spine

1. Perception

Build and diagnose visual signals across segmentation, matting, depth, detection, tracking, and multimodal inputs.

2. Reasoning

Turn model outputs and product intent into explicit state, rules, prompts, schemas, and controllable workflows.

3. Runtime

Connect models to C++/Android, GPU pipelines, services, APIs, provider routing, and deployment constraints.

4. Evaluation

Make quality inspectable through hard cases, comparison reports, traces, failure taxonomies, and release gates.

5. Handoff

Package assumptions, interfaces, debug paths, acceptance criteria, and operating knowledge for the next team.

Operating Principles

  • Workflow first. A useful AI system starts with the product or operational problem.
  • Deployment is modeling. Architecture, I/O, quantization, runtime, and debug tooling are one design space.
  • Edge cases decide quality. Camera AI is won in boundaries, occlusion, motion, lighting, and device variation.
  • Evaluation before claims. Multimodal outputs should be structured, compared, reproduced, and debugged.
  • Public-safe by default. I share reusable judgment without customer names, internal paths, artifacts, logs, or private metrics.