I turn frontier model capabilities into products people can actually use.

The problem

Running a swarm of Claude Code agents in parallel is powerful but illegible. Output streams interleave in a flat terminal, work-in-progress is invisible until something breaks, and there is no way to compare what 5 agents are doing against the same plan.

The constraint

28 hours, solo, at the SensAI hackathon — cross-platform spatial UI from scratch.

The move

I built VibeView as a WebSpatial app that arranges live Claude Code sessions as panels you can walk around. Each agent gets its own surface streaming stdout/stderr; a shared planning layer sits above them; dependencies between agents render as spatial links rather than text. The hard call was choosing WebSpatial — it ships to visionOS, web, and mobile from one codebase, but the docs are sparse, so I had to reverse-engineer the renderer from runtime behavior. I kept the orchestration logic in a thin Node bridge so the spatial layer stayed presentational.

The outcome

Shipped end-to-end in 28 hours, demoed live driving a 5-agent swarm against a real codebase. Won a category prize and a follow-up call from the WebSpatial team.

The problem

Engineering managers lose 30–60 minutes per meeting translating decisions into tickets, and the translation is lossy. Action items slip, ownership is ambiguous, and the trail back to the conversation that produced them is gone.

The constraint

Solo founder, B2B SaaS, in active enterprise sales cycle — must hit security review on day one.

The move

Aligned Tools ingests a transcript, extracts decisions and commitments with a Claude pipeline, and writes them straight into Jira with parent/epic links and the relevant transcript span attached. The product surface is intentionally narrow: paste or upload, review the proposed tickets in a diff-style UI, push. I designed the review UI so the human never has to type — every edit is a one-click correction that becomes training signal. Auth is SSO-only; data is tenant-isolated; transcripts are dropped after ticketization unless the customer opts in.

The outcome

Closed first paid pilots with two engineering orgs; in active procurement at a Fortune 500. Median time-from-meeting-to-tickets dropped from ~45 min to under 3.

The problem

Claude Code's power lives in env vars, hooks, slash commands, and skills — but they are scattered across files, undocumented in-product, and impossible to compare across projects. New users hit a wall; power users build the same scaffolding over and over.

The constraint

Consumer-facing desktop app, must feel native on macOS, ship a marketplace from day one.

The move

SkillVault Desktop is a single pane that reads your Claude Code config across every project, surfaces what is actually wired up, and lets you install vetted skills/hooks/commands from a marketplace with one click. The harder design call was the diff: when a user installs a skill, they see exactly which files change, with rollback. I built it on Tauri so it stays a 12 MB binary instead of a 200 MB Electron app, and the marketplace is just signed manifests in a public repo so anyone can publish without a backend.

The outcome

Public beta launched with 40+ community-contributed skills in the first week. Average user activates 6 skills in their first session.