Work

Problem. Once you run several coding agents in parallel, the bottleneck stops being code and becomes context. Close a terminal, lose the session, and the agent forgets everything about your project. Traditional IDEs bolt a chat panel onto a human-first workflow.

What it is. An engineering environment designed for agents from the ground up: multiple AI coding agents managed together, persistent sessions that can be resumed, pipeline-driven workflows with automated handoffs, an extensible architecture for future tools — and humans holding the engineering decisions.

Status. Open-sourced in 2026 and being built in public.

Problem. Travelers across Asia share their trips constantly, but the tools they use were not designed around sharing as the primary act, and rarely with English-speaking users in the region in mind.

What I’m doing. Designing TripPro around social sharing × social tech from the start: how a trip gets captured, how it gets shared, and how it reaches the people it’s meant for. Development is demo-driven — the working demo is the source of truth for the product direction.

Status. In progress. I’m writing about the decisions as I go.

Problem. Driving a normal browser with an agent means the agent and the human fight over the same tabs and windows, or the agent runs in a separate profile with none of the user’s sessions.

What I did. Worked on a Chromium-based browser designed for both people and agents. Agents get their own isolated task space, reuse the user’s existing login state, and clean up after themselves when a task completes. Along the way we diagnosed and fixed renderer-process accumulation caused by per-domain site isolation.

Result. Used day to day for browser automation and web testing; a write-up of the design and the renderer investigation is coming to the blog.

Problem. Users want to paste a social profile or post link and get useful data back. Several platforms — Instagram in particular — have moved their internal APIs behind a login wall.

What I did. Built an MVP that falls back to what remains publicly available: Open Graph metadata, fetched with crawler user-agents that platforms still serve previews to. Handled per-platform differences and kept the approach within robots.txt and terms-of-service boundaries.

Result. A working import path without official APIs, and a clear map of what each platform still exposes.