I build AI agent product workflows: tool integrations, local-first runtimes, multimodal artifacts, Agent UX, and human-in-the-loop automation.

idea -> tools + models + files -> reviewable workflow -> product capability

Most of my work is hands-on: real product code, clear engineering boundaries, and workflows that move from a rough idea to something users can run, review, recover, and reuse.

AI agents are becoming useful, but the hard product work sits around the model: tools, vendors, files, media, local runtimes, permissions, progress states, error recovery, review loops, packaging, and operational checks.

That is the layer I like building.

These are the public projects that best show how I think about agents: not as chat boxes, but as product systems with tools, state, artifacts, review loops, and operational boundaries.

• Tool-integrated agents: external tools, vendor APIs, auth, schemas, async jobs, output handling, and failure boundaries.
• Local-first runtimes: desktop sidecars, CLIs, filesystem state, packaged apps, and reproducible workspaces.
• Multimodal workflows: files, images, audio, video, documents, previews, exports, limits, and validation.
• Agent UX: visible state, understandable errors, cancellation, retry, recovery paths, review gates, and human control.
• Workflow productization: turning ambiguous ideas into stable, testable, repeatable product flows.
• Commercialization design: shaping AI capabilities into usable product surfaces, upgrade paths, and operational workflows.

I am shaping Caprika Agent Lab as a proof-of-work collection for small, runnable AI product patterns.

The goal is to show the path from idea to product capability, not just the final screenshot.

Planned examples:

• tool wrappers with clear permission and error boundaries
• multimodal input/output flows for files, images, video, and documents
• content workflows with source evidence, human review, and publish/readback loops
• Agent UX patterns for progress, cancellation, retry, and recovery
• local runtime patterns with CLIs, daemons, filesystem artifacts, and previews

• I prefer inspectable artifacts over hidden state.
• I like runnable slices that prove a capability before scaling it.
• I care about agent systems that survive real repositories, real users, code review, packaging, deployment, and maintenance.
• I treat product, engineering, and operations as one system when agents are part of the workflow.

• GitHub: @alchemistklk
• Location: Shanghai
• Focus: AI agent products, local-first tools, multimodal workflows, and tool-integrated automation