Joseph
Mattiello

The challenge

Video assets across product surfaces were being delivered with no shared compression or caching strategy. At Wayfair's scale, bandwidth costs grow fast — and the cost was hidden across teams, so no one owned it.

What I did

• MP4 pipeline rewrite. Profiled the existing video delivery path, designed a re-encoding + caching strategy, and validated the savings against real production traffic via Datadog RUM.
• SwiftUI purchase flow profiling. Instrumented the highest-revenue purchasing flow, identified rendering hot paths, and shipped targeted SwiftUI optimizations.
• Bootup performance instrumentation. Stood up Datadog RUM tracking for cold-start, splash-to-interactive, and warm-resume timings — the data that's now driving the next wave of perf work.
• AI tooling for the platform team. Built custom MCP servers, Claude skills/plugins, and n8n workflows that integrate with Jira and GitHub to automate cross-team engineering tasks.

Outcome

Roughly $2M/year in recurring bandwidth savings from the video work, real bootup metrics now feeding a structured optimization roadmap, and a growing internal AI tooling surface used across teams.

The challenge

iOS apps across NewsCorp's brand portfolio were lagging Android on app-store ratings, crash rates, performance, and bandwidth. Each brand was duplicating platform code. CI/CD was inconsistent. There was no shared iOS strategy.

What I did

• Built a shared iOS Platform SDK as an "internal open source" model — every NewsCorp brand consumed it, contributed back, and got the next perf/stability fix for free.
• Restructured and hired across 8 regions. Set up cross-region code review, release cadence, and on-call ownership.
• Modernized CI/CD with GitHub Actions, automated PR management, and Jira integrations — the same DevOps automation patterns I'd later carry into my OSS work.
• Shipped Knewz from scratch — NewsCorp's news-aggregation app — as a greenfield iOS project, end-to-end, while running platform.

Outcome

iOS surpassed Android across the board on stability, app-store ratings, performance, and bandwidth usage. The platform SDK was adopted by every brand and outlived my tenure. Knewz launched on iOS internally-built, on time.

The challenge

The React Native iOS build was slow and flaky — both in CI and locally. Engineers had stopped trusting green builds. And video quality was capped by what JavaScript could push through the bridge.

What I did

• Rebuilt the iOS RN build pipeline. Profiled every stage, eliminated flakiness sources, and parallelized where it mattered. CI went from "maybe" to predictable.
• Shipped HD video as a native React Native module. Built compression, transcoding, and playback in Swift/AVFoundation, exposed via the React Native bridge — gave product the quality bar they couldn't get from JS.
• AR filter exploration. Prototyped native AR filters using ARKit / Core Image, demonstrating the headroom available once you stop fighting the bridge.
• CI/CD migration to streamline the team's release cycle.

Outcome

Predictable CI, HD video shipped natively, and a clear pattern for the team: when the JS bridge isn't enough, here's how we drop down to native. Engineering velocity recovered.

The challenge

Maintaining a 13-year-old open-source project with 6.3k stars, 79 contributors, 388k+ downloads, and an active App Store presence — while working full-time. Solo-maintainer velocity is the rate-limiter. Generic "use Copilot" doesn't move the needle.

What I built

• Reviewer-model auto-selection by static diff analysis. The PR review workflow grep's the diff for thread-safety patterns (Realm, @Model, actor, @MainActor) — simple PRs route to Claude Sonnet; thread-safety-sensitive PRs route to Claude Opus. Real cost optimization, not vibes.
• Three agents, one action vocabulary. Claude Code, Cursor Agent, and Kimi Code share dispatch verbs — implement_issue, fix_ai_review, fix_rebase_conflict, ai_approved — routed by branch name and PR labels. Effectively a multi-agent orchestration layer on top of workflow_dispatch.
• Kimi Code as Anthropic-API-compatible fallback. When Claude credits run low, the same Claude Code CLI binary redirects via ANTHROPIC_BASE_URL to drive Kimi — zero workflow rewrite.
• Scheduled Copilot review poller. GitHub blocks event-triggered workflows from bot actors. The 10-minute cron poller bypasses the approval gate by running as a trusted scheduled actor — a deliberately documented workaround.
• Auto-fix CI on the wiki repo. Spell-check / link-check failures spawn Claude via anthropics/claude-code-action, which fixes the root cause and opens its own PR (with Copilot reviewer auto-requested).
• AI-driven localization. Diff-aware Claude Haiku pipeline translates only changed strings into 11 languages — cost-controlled by extracting just the new keys, not the whole file.
• Repo health monitor. Cron job detects stuck PRs, re-dispatches reviews, and maintains a pinned automation-health issue.
• Upstream-merged Claude Code workflow. Installed Claude Code into the libretro upstream itself (PR #107) — alongside 19 merged feature PRs covering save states, RetroAchievements, SIMD blitter, WASM/Emscripten, and Jaguar CD support.

Outcome

OSS velocity sustained at scale on a solo-maintainer schedule. The pipeline handles routine review, rebase conflicts, localization, and CI failures autonomously — escalating to me only when bounded review cycles complete or 3-cycle limits are hit. This is what "AI engineering leadership" looks like when it's real, not slideware.

Proof — read the code

• ai-review.yml — Sonnet vs Opus auto-selection by static diff analysis
• cursor-agent.yml + kimi-agent.yml — three agents, shared dispatch vocabulary
• copilot-review-poller.yml — bot-actor gate bypass via scheduled cron
• auto-translate.yml — diff-aware Claude Haiku localization (11 languages)
• wiki/auto-fix-ci.yml — Claude self-healing on CI failures
• libretro PR #107 — Claude Code installed in upstream libretro

The challenge

Bare-metal Atari Jaguar is the kind of hardware LLMs don't know — released 1993, niche m68k+RISC architecture, ~half a dozen active homebrew developers worldwide, a cross-compiler that hasn't seen a release in 20 years. Conventional wisdom says "AI can't help here." This case study disproves that.

What I built

• Three-agent ship→review→address-feedback loop captured in the commit log. Examples: cd97ceb "review: address Copilot/Qodo feedback on PR #5"; PR #25 entirely dedicated to addressing Copilot review on merged PR #24. The iteration is in the git history, not just the chat transcripts.
• Cross-cutting hardware-aware change. PR #27 (PAL pass) — region-override toggle in Options menu, region-aware geometry replacing hardcoded H=240 with PALNTSC ? 240 : 288 across 10 procedural pattern functions, 50 Hz timing for AudioSync + MDFourier, plus VDP register diagnostics. Exactly the type of cross-cutting refactor LLMs typically struggle on bare-metal.
• 7 hardware probe features shipped. 93C46 EEPROM read/write, Jaguar CD/Butch probe, JagLink serial test, Memory Track EEPROM, Sprite Stress / OP saturation, Pro Controller + Rotary controller, Video Mode TOM register probe.
• Custom Jaguar SDK in a Docker image, published to GHCR. The .github/workflows/sdk-image.yml builds ghcr.io/joematt/jaguar-sdk (m68k-atari-mint-gcc, mac, aln, jlibc/rmvlib) so build/release can pull it instead of recreating the toolchain. Solves the "exotic cross-compile toolchain" problem by containerizing it.
• Image-diff-style smoke-test CI. .github/workflows/build.yml's smoke-test job checks out libretro/virtualjaguar-libretro upstream, caches the built core keyed on resolved upstream SHA, boots the freshly built ROM via a Python libretro.py harness for 60 frames, dumps the framebuffer as a PNG artifact, and verifies fastboot signature byte-identical to upstream. Real visual regression testing for a ROM project.
• Visual regression harness. PR #35 / commit 0fc3eb3 expands a scripted screenshot tour from 19 → 47 navigable test screens, captured automatically via scripts/screenshot-tour.py. Closest thing the bare-metal codebase has to visual diffing.
• Disciplined release engineering. 12 tagged releases (v0.6.4 → v1.3.1) in 10 days. Each feature PR followed by a dedicated release: bump version to X.Y.Z PR.

Outcome

Goes against the prevailing narrative that AI-assisted development only works on well-trodden stacks. Bare-metal Jaguar is the test case where it absolutely shouldn't work — and the git history shows it does. The pattern is the durable insight: AI as primary author + adversarial reviewers + iteration captured in commit history.

Proof — read the code

• PR #27 — PAL pass, region-aware geometry across 10 pattern functions
• PR #25 — entirely addresses Copilot review of merged PR #24
• PR #35 — Screenshot tour expanded to 47 screens
• build.yml — smoke-test CI booting every PR ROM
• sdk-image.yml — Custom Jaguar SDK Docker image
• Releases — 12 releases in 10 days