Senior Frontend Engineer & Agent Systems Architect.

• 01
  Case 01Sunflower Games2025 – 20267 months

  Read case

  Roulette mini-game, missions, and ops tooling.

  Ops and gameplay teams needed better tools while a live dual-currency economy kept moving underneath them.

• 02
  Case 02FeatherShark2023 – 202420 months · 2-person team

  Read case

  Live govtech migration, no release gaps.

  The app needed a new frontend foundation, but municipal users depended on weekly releases that could not pause.

• 03
  Case 03MongoDB2023 – 202418 months

  Read case

  Design-system fixes on mongodb.com.

  High-traffic marketing pages needed library improvements, but dozens of consumers depended on those packages.

• 04
  Case 04Platform92018 – 20234+ years

  Read case

  Shared UI for a private-cloud console.

  The console needed reusable UI building blocks and a predictable way to cache and persist server data, instead of components and fetch logic re-solved case by case.

Tools I build and run daily for agent orchestration, agent memory, and job-search automation. Each one exists because I hit the problem myself and no existing tool solved it.

• RepoKernel

  Agent orchestration in isolated Git worktrees.

  Spec-first sprints for coding agents: isolated worktrees, small task scopes, controlled context, dependency ordering. Nothing merges without passing a review gate.

  • TypeScript
  • AI agents
  • Git worktrees
  • DevTools

  github.com/xantorres/repokernel

• Engram

  Local-first memory for coding agents.

  Captures facts from any coding agent, gates sensitive writes behind a review queue, and recalls them across tools. Agent-agnostic and MCP-native.

  • Python
  • MCP
  • Local-first
  • Agent memory

  github.com/xantorres/engram

• Shrike

  Local-first CRM for serious job search.

  Ingests, filters, scores, and tracks job opportunities with AI-assisted triage and hard rejection rules. I ran my own search on it.

  • TypeScript
  • CLI
  • AI triage
  • Job search

  github.com/xantorres/shrike

• 01

  Architecture that holds up

  I map data flow, state ownership, and failure paths before the first component lands. Boring on purpose: fewer rewrites, fewer surprises, and code the next engineer can navigate without a guide.

• 02

  Frontend craft with receipts

  Interaction detail, Core Web Vitals, accessibility, loading states, and the edge cases users always find. On mongodb.com that meant a two-line fetchPriority change with a measurable LCP win.

• 03

  Product judgment

  Comfortable with vague requirements, stakeholder pressure, and incomplete information. I argue for the version users need, cut scope honestly, and delete code that duplicates truth owned elsewhere.

• 04

  Codebases teams can live with

  Typed boundaries, migrations that never freeze feature work, and CI that catches regressions before main. Six stack migrations on one live govtech product without a single release gap.

1. 01

   Map the system first.

   Before touching components I want to know where data comes from, who owns state, and what breaks under failure. An hour of mapping saves a week of rework.

2. 02

   Use AI without losing control.

   AI speeds up implementation, exploration, and refactors. Architecture, review, testing, and product decisions stay human-owned, so velocity never outruns judgment.

3. 03

   Trust one source of truth.

   More than once, I have deleted a client-side recomputation of a backend value and made one endpoint authoritative instead. Each time, an entire class of drift bugs disappeared with it.

4. 04

   Ship the useful version.

   I work async and remote by default. I make clear calls, write down the tradeoffs, and keep momentum when waiting for perfect consensus would stall the work.