Solutions Architect · AI & Full-Stack Engineering
Get in touch ·
Explore my projects
Rea · 458K+ stars
Original PR #7465 · Merged upstream via #10406
Fixed misleading MCP Doctor diagnostics when browser or native tool scopes skip the identity probe. The diagnostic now explains why identity was not observed without changing the health check or exit status. Added regression tests for both scopes. The fix landed on main through #10406 with authorship preserved.
Docling · 70k+ stars
Contributed to a document-processing toolkit for generative AI, with integrations across LangChain, LlamaIndex, and the broader AI ecosystem.
Merged PR #5228 — Prevent silent content loss in HTML email ingestion
Fixed silent loss of HTML email footers by preserving content-layer semantics end to end, from the HTML backend through the email document model and JSON export. Footer content is also available through opt-in Markdown export, while default body-only output and plain-text precedence remain unchanged.
Added end-to-end regression coverage for HTML-only and multipart messages, blank plain-text alternatives, footer-only content, escaping, and consistency across the document model and export formats.
Karakeep · 30k+ stars
Contributed to an open-source bookmarking app with AI-powered tagging and full-text search.
Merged PR #3053 — Search by notes and highlights
Implemented has:notes and has:highlights filters across database queries, smart lists, and search autocomplete, with documentation and regression tests.
Merged PR #3299 — Consistent manual summary status
Updated successful manual summaries and their status in the same database write, with regression tests for recovery, inference errors, empty responses, and cross-user access.
SayAll · 2.6k+ stars
Contributed to a native macOS app that turns Bluetooth voice remotes into wireless microphones and customizable controls.
Merged PRs: #829 · #844 · #888 · #913
Fixed left/right modifier-key handling, unified native menu-bar interactions, added per-device remote battery status, and made focus-option cards fully clickable. Included regression tests and manual macOS validation.
| Project | What to look for | Stack |
|---|---|---|
| Agent Lens | Compare report quality across prompt versions. Inspect human review, judge calibration, and cost/latency tradeoffs. | Python · FastAPI · React |
| Signal Workbench | Run an offline research pipeline with synthetic crypto evidence, proposal validation, budget fallback, and resumable local delivery. | Python · SQLite |
| Reply Copilot | AI writing assistance inside the browser, with reviewable context, explicit site access, and protection against overwriting newer drafts. | JavaScript · Chrome MV3 |
| CloseLoop | A decision-management PWA with explicit state transitions, preserved reasoning, and separate data/asset caching. | TypeScript · React · D1 |
Also building Agent Title Renamer: a portable skill for keeping AI conversation titles useful and consistent.
Away from the keyboard, I share my home with four cats. They also supervise the keyboard. (´•༝•)`
- Make behavior inspectable. Small modules, documented decisions, and examples you can run locally.
- Test the failure paths. Invalid model output, stale editor state, interrupted delivery, and data boundaries deserve explicit handling.
- Connect architecture to delivery. System boundaries, integration choices, APIs, and data pipelines through to the interface people use.
Tools I reach for: Python, FastAPI, SQLAlchemy, SQLite, TypeScript, React, Chrome extensions, Cloudflare Workers, D1, and GitHub Actions.
Open to remote opportunities in solutions architecture, AI applications, research tooling, data automation, and full-stack development. Comfortable with asynchronous collaboration in English and Mandarin.



