AI orchestration developer

213 merged_PRs

contributions

days

grazomarin

AI orchestration developer building adaptive systems, agent workflows, and thoughtful interfaces.

Background visual: a trace of orchestration runs — a feature shipped through a six-way subagent fan-out, context assembly, cross-model review, model judging, memory distillation.

> git log --since=2022

Build Log

  • The Odin Era (2022): Where it started — a year of vanilla JavaScript with no frameworks and no shortcuts.
    • Battleship (2022): A test-driven Battleship built with Jest — no framework.
    • Knight's Travails (2022): Shortest path for a knight via breadth-first search.
    • Weather App (2022): A weather app on raw fetch + async state.
    • Calculator (2022): An on-screen calculator handling operator chaining and edge cases.
    • Library (2022): A book-tracking app built around a constructor-based object model.
    • Tic-Tac-Toe (2022): Tic-tac-toe structured with the module/factory pattern.
    • Etch-a-Sketch (2022): A dynamic hover-drawing grid with adjustable resolution.
    • Memory Card (2022): A memory game scoring unique clicks against fetched card art.
    • CV Project (2022): A CV builder that renders live from form input.
    • Restaurant Page (2022): A tabbed restaurant page assembled entirely in JavaScript via webpack.
    • To-Do List (2022): A to-do app with projects, priorities, and localStorage persistence.
    • Form Validation (2022): A form with live, constraint-based client-side validation.
    • Dashboard (2022): An admin dashboard layout built to practice CSS Grid.
    • Sign-up Form (2022): A styled sign-up form with matching-password validation.
    • Rock Paper Scissors (2022): A five-round rock-paper-scissors game played in the DOM.
    • Odin Recipes (2022): A static recipe site — the first multi-page HTML build.
  • UBC Era (2022–24): University: CS fundamentals, the first AI experiments, and accessibility groundwork.
    • BST Visualizer (2023): A binary-search-tree visualizer rebuilt in TypeScript.
    • AI Wardrobe (2024): First AI-powered build — prompt-driven outfit logic; the moment the path bent toward AI.
    • WCAG Guidelines (2024): A WCAG project — the a11y groundwork that resurfaces in the DatePicker.
    • UBC AI Club (2024): Work with UBC's AI club — a club site and applied-AI workshops.
    • Real-Estate Regression (2024): A regression model predicting housing prices from feature data.
    • JavaFit (2023): A Java fitness-tracking app — CS-fundamentals coursework in OOP.
    • Grade Calculator (2023): A weighted-grade calculator for tracking course standing.
    • VS Code SPD Handin (2024): Tooling around the VS Code assignment-handin workflow.
  • Samsung (2025–26): Enterprise front-end at scale, plus the AI-infrastructure work — Graphify and Memory Bank.
    • Build & Quality Infra (2025): Performance, CI parity, and stability across enterprise front-ends at Samsung R&D.
    • Component Library (2025–26): Design-system components and typed utilities for Samsung's shared library.
    • DatePicker System (2026): A production DatePicker that survives edge cases: segment input, a full keyboard model, timezone correctness, localization.
    • Graphify Integration (2026): Adapted the open-source Graphify knowledge-graph tool to a large Samsung monorepo — durable structural memory for an AI assistant.
    • Memory Bank (2026): A two-tier context system turning a stateless AI assistant into a persistent collaborator, for Samsung's Cline fork.
  • ÖzünÖyrən (2026): Lead full-stack on an LMS platform — admin frontend, Spring Boot backend, the engineering substrate.
    • Admin Course Platform (2026): Owned the LMS admin frontend — the full /admin/courses epic on Next.js 16 / React 19 / Chakra UI v3.
    • Management API (2026): Drove a large slice of the Spring Boot management backend — the course/educator contract the frontend ran on.
    • OIDC / Keycloak Auth (2026): Implemented Keycloak OIDC auth for admin-ui and migrated it to a centralized dev auth server.
    • Contract Harness (2026): Built a net-zero contract-verification harness checking admin-ui's assumed API against the live dev backend.
    • Engineering Conventions (2026): Owned the repo's contribution conventions and the CI/hook machinery that enforces them (ADR5).
  • AI Tooling (2024–26): The cross-project methodology that travels with me — review pipelines, skills, and a memory architecture.
    • Parallel PR Review (2025–26): A multi-agent review pipeline: ~6 specialist subagents fan out, then a strict scrutiny pass filters every finding.
    • Split-PR (2025): A skill that decomposes a large PR into a stack of smaller reviewable PRs with a byte-identical tree guarantee.
    • Dual Review (2025): A sequential cross-model review of design artifacts — round 1 Codex (GPT-family), round 2 a fresh Claude reviewer.
    • Session-End Ritual (2026): A session-end consolidation ritual that distills each session's durable lessons into the persistent memory layer.
    • Verification Guard (2025): A verification-guard harness whose blank-context verifiers RAN the code before allowing a commit — built, validated, then deliberately retired.
    • Memory Architecture (2025–26): A hybrid institutional-memory system for AI sessions: a cross-project rule layer symlinked into each repo's local memory.
directthematic

> git log --since=2022

The Through-Line

From a test-driven Battleship in vanilla JS to multi-agent AI orchestration — the origin is where it started; each arc outward is a year. Open a chapter to drill into the work.

how to read

· distance from origin = time

· each node is a chapter — click to open

· ── direct lineage   ╌╌ thematic thread

2022–2026 · 5 chapters · ~40 projects · vanilla JS → AI

Build Log overview

> man orchestrate_intelligence

The loop

The trace in the hero is not a mock; it replays how I actually ship. By day I build front-end systems at Samsung R&D. The rest, the agents, the memory, the review pipelines, exists because I wanted shipping with AI to feel as disciplined as shipping without it.

Every run walks the same six stages.

assemble_context()

A run starts by remembering, and it happens on its own. The harness loads the distilled rules of past sessions, then the repo itself: structure, history, the graph of what touched what. No run starts from zero.

  • curated memory · 30+ distilled rules
  • Graphify graph
  1. assemble_context()

    A run starts by remembering, and it happens on its own. The harness loads the distilled rules of past sessions, then the repo itself: structure, history, the graph of what touched what. No run starts from zero.

    • curated memory · 30+ distilled rules
    • Graphify graph
  2. forge_spec()

    An idea is brainstormed into a spec before any code exists. Then the spec gets grilled: question after question until every branch of the design tree is resolved. What survives is worth building.

    • superpowers:brainstorming
    • /grill-me
  3. write_plan()

    The spec becomes a plan, the plan becomes tasks small enough to review honestly. Before anything runs, the plan faces three different models, and each must sign off.

    • writing-plans
    • plan review: claude · codex · gemini
  4. fan_out()

    Implementation goes to fresh subagents, one per task, each handed exactly the context it needs and nothing more. Isolation keeps them focused; the orchestrator keeps the thread.

    • subagent-driven development
    • fresh context per task
  5. scrutinize()

    Every change faces reviewers that did not write it: spec compliance first, then code quality, then a second model with different instincts. A finding only counts once it is verified against the source. Then the polish pass lands.

    • parallel-pr-review
    • dual-review: codex → claude
    • judge_panel: gemini · codex · claude
  6. distill()

    /capture ends the session: what happened is distilled into memory. Corrections become durable rules, overlapping rules merge, and the next run assembles a smarter context than this one had.

    • /capture
    • write-back → memory index

> open_channel()

Send a signal

If you're hiring for this kind of work, or building with agents and want to compare notes, the channel is open. Email gets a reply; GitHub is where the practice lives in public.

session persisted · memory updated · exit 0