Wilfredo Baez, AI product engineer in New York
I'm an AI product engineer in New York. I build products end to end by directing AI coding agents, and I test what ships. Open to founding and early engineering roles.
Work
-
236gate tests
For the Record iOS music diary236 gate tests
iOS app · In App Store review
For the Record
A music diary for every genre: rate albums in half stars, log listens, keep a crate and follow people whose taste you trust.
- Problem
- Film fans have a place to log what they watch. Music fans who want the same thing for albums have few native options that cover every genre and sync cleanly across devices.
- My role
- Solo developer and publisher. I built the app, the sync engine, the backend schema and its security rules, the legal site, the App Store listing and the release pipeline.
- Shipped
- Version 1.0 of the iOS app, submitted through App Store Connect, plus its backend and a marketing and legal site.
- Stack
- Swift
- SwiftUI
- Supabase (Postgres, row-level security, edge functions)
- Sign in with Apple
- StoreKit 2
- MusicKit
- Outcome
- 236 gate tests green before release (127 app core, 109 backend)
- A release check that fails any build calling an OS symbol the target iPhone lacks, written after it caught a real crash
- Build uploaded to App Store Connect in November 2026
-
156items synced in one minute
Lagoon Private visual board for iPhone and iPad156 items synced in 1 min
iOS app · Preparing for the App Store
Lagoon
A private visual board for iPhone and iPad. Drop in photos, videos, links and notes from any app. No AI, no account.
- Problem
- Visual bookmarking apps upload everything to their own cloud and keep adding AI features that many people do not want.
- My role
- Solo developer: app, share extension, storage format, iCloud sync, migration and recovery tools, paywall and policy pages.
- Shipped
- iPhone and iPad app with a share extension and an export and import backup format.
- Stack
- Swift
- SwiftUI
- Liquid Glass (iOS 26)
- iCloud Drive sync
- StoreKit 2
- AVFoundation
- Cloudflare Pages
- Outcome
- Sync stores every item as its own file with tombstones for deletes, so two devices never overwrite each other
- Verified on three devices: a fresh iPad went from 0 to 156 synced items about a minute after launch
- App Store submission planned for November 2026
-
1,300+tests across 52 services
Vigil A personal agent system52 services, 1,300+ tests
Agent system · Runs every day
Vigil
A personal operations system that scans every project, plans the day, runs a crew of role agents and leaves every irreversible step to a human tap.
- Problem
- Running a client business alongside many builds meant no single place knew what was open, what was owed and what could safely be done without me.
- My role
- Designer and builder: architecture, agent roles, guard hooks, the API contract, the approval model and the native apps.
- Shipped
- A CLI, an always-on local server, a dashboard and native apps, plus a build department (architect, coder, tester, reviewer) that runs jobs in parallel git worktrees.
- Stack
- Python
- SQLite
- Claude Code CLI
- Claude Code sandbox and guard hooks
- SwiftUI (iPhone, iPad, Mac, Watch)
- APNs
- Notion API
- launchd
- Outcome
- 52 services behind one versioned HTTP contract
- About 1,300 Python and 270 Swift test functions, plus 28 eval scripts
- Agents draft and prepare; sending, posting, paying and deleting always wait for my approval
-
28/28planner eval passing
Sundial Day planner with a local AI co-plannerPlanner eval 28/28
iPhone and Mac app · In daily use
Sundial
A visual day planner with a local AI co-planner that drafts your day from your real tasks.
- Problem
- Visual planners do not connect to the task systems people already use, and they send planning to the cloud.
- My role
- Solo developer: the app, CloudKit sync and the Mac-side planner service.
- Shipped
- iPhone app and Mac app sharing one CloudKit store.
- Stack
- SwiftUI
- SwiftData
- CloudKit
- EventKit
- Python planner service
- Claude Code CLI
- Outcome
- Planner eval passes 28 of 28 cases
- Mac to iPhone sync verified with matching row counts in both stores
- Planning runs on my own Mac, not a third-party API
-
66/66moves animated, drawn in code
Cadence Daily calisthenics, demos drawn in code66 of 66 moves animated
iPad app · In daily use
Cadence
A 15-minute daily calisthenics plan with an animated demo for every move, drawn in code.
- Problem
- Bodyweight plans need a clear demo for every exercise, and shipping a video library makes the app heavy and hard to change.
- My role
- Developer: the app and a small animation engine that poses a stick figure with inverse kinematics.
- Shipped
- iPad app, with iPhone planned.
- Stack
- Swift
- SwiftUI
- HealthKit
- Live Activities
- Swift package for the figure engine
- Outcome
- All 66 built-in moves have an animated demo
- Gate tests fail the build if any exercise lacks a demo, a limb changes length or the figure drops below the floor
-
In useat a client practice
Clinical AI Documentation Notes tooling for a therapy practicePrivate client work · Mirrored tests on the real form
Client tool, web and iOS · In use at a client practice
Clinical AI Documentation
A note-writing tool for a therapy practice that fills the clinician's existing records system, so nobody changes how they work.
- Problem
- Clinicians spent a large part of each day writing session notes and treatment plans by hand into their records system.
- My role
- Sole developer for the client: requirements, integration with the practice's records workflow, sensitive-data handling, deployment and support.
- Shipped
- A web app plus a desktop bookmarklet and an iOS Safari extension that fill the records system's forms.
- Stack
- Node.js
- Express
- Postgres
- LLM note drafting
- Browser bookmarklet
- iOS Safari extension
- JSDOM tests
- Outcome
- The desktop and iOS fillers are held to the same behavior by mirrored test suites
- Every fill is built against a captured copy of the real form, not a simplified one
- Client name, data and screens stay private by design
-
844open listings, Nov 2026
Hunt Board Job and grant board844 listings, Nov 2026
Web tool · Public version live
Hunt Board
A job and grant board that pulls live listings, scores them against a resume, and checks every tailored line against what the resume actually says.
- Problem
- Listings are spread across dozens of sites, and AI resume tailoring tends to make up experience.
- My role
- Solo builder: feed integrations, scoring, the application tracker, tailoring with a fabrication guard, a Mac wrapper app and a health check.
- Shipped
- A private dashboard, Mac app and CLI for my own search, plus a public board anyone can use with no sign-up.
- Stack
- Python standard library
- SQLite
- RSS and public JSON APIs
- Claude Code CLI
- Swift WKWebView app
- Outcome
- Public board: 844 open listings from 6 sources whose terms allow it, live since November 2026
- Private board: 11 live feeds and 22 one-click searches
- Private board: 214 tests in the gate
- Four eval suites run on real model calls: resume parsing, tailoring, job fit and picks
- November 2026 runs scored 96 to 100 percent, each suite against its own pass bar
- The fabrication guard closed a real leak where stray phone digits let an invented "10 years" through
-
474items per second, purity 0.93
TopicLens On-device topic sorting474 items/s, purity 0.93
On-device machine learning · Shipped for personal use
TopicLens
Sorts anything you have read or written into topics, live, on the Mac's Neural Engine. Nothing leaves the machine.
- Problem
- Big personal text collections are hard to browse by theme without sending them to a cloud model.
- My role
- Builder: clustering, topic naming and search on top of an on-device embedding model, plus the Mac app, CLI and eval harness.
- Shipped
- Native Mac app and CLI.
- Stack
- Swift
- SwiftUI
- Core ML
- EmbeddingGemma on the Apple Neural Engine
- Outcome
- About 474 items per second end to end on an M4
- Topic purity 0.93 on 10,080 labelled posts
- Search precision at 10 of 0.85
-
0.2%CPU, down from 12-25%
Claude Pulse A Mac widget for live AI sessionsCPU 12-25% → 0.2%
Mac developer tool · Shipped for personal use
Claude Pulse
A floating Mac widget that shows what every AI coding session and build job is doing right now.
- Problem
- Running many agent sessions at once meant clicking around all day to see which were working, stuck or done.
- My role
- Solo builder: the session-state collector, the state rules, the native widget and the CPU tuning.
- Shipped
- A floating Mac app and the background collector that feeds it.
- Stack
- Python
- SwiftUI
- Core Animation
- Versioned JSON contract
- Outcome
- Cut the widget's CPU use from 12 to 25 percent down to about 0.2 percent by moving animation to Core Animation layers
- An audit script re-checks a week of sessions against the state rules
-
104tests; code checks every write
Brain Knowledge engine where code checks every write104 tests
Knowledge engine · In daily use
Brain
A personal knowledge engine where the model only proposes and code checks and writes, rendered into an Obsidian vault.
- Problem
- AI note systems let the model write freely, which lets invented facts and broken links into your notes.
- My role
- Solo builder: ingestion, validation rules, a privacy filter, the Obsidian export and daily-note capture.
- Shipped
- A CLI engine and the vault it writes.
- Stack
- Python
- Claude Code CLI
- Obsidian (Markdown and Bases)
- git
- Outcome
- 104 tests pass
- Evals run against an invented owner, so the test data holds nothing personal
- A privacy test fails on any excluded term
-
No edgemeasured after fees, stated plainly
Cash Cow and TradeDesk Paper trading researchNo edge after fees
Market research, paper only · Paper only, no real money
Cash Cow and TradeDesk
A research terminal where code decides which price moves matter and the model only explains them, plus a paper desk that refuses to claim an edge it cannot show.
- Problem
- AI market commentary mixes invented reasons with real ones, and strategy dashboards show flattering numbers without fees.
- My role
- Solo builder: price feeds, rules, a backtester with a held-out period, paper strategy books and a native Mac dashboard.
- Shipped
- A CLI research terminal, a paper trading engine and a Mac dashboard.
- Stack
- Python
- Public price data
- Claude Code CLI
- SwiftUI Mac app
- launchd
- Outcome
- Results reported as they came: no paper book beat buy and hold after fees over 24 months
- An eval fails the build if the dashboard claims an edge that does not survive the checks
- There is no live trading mode
-
50/50parity checks
Clean-room pipeline Clean-room rebuilds with a leak gate50/50 parity checks
Developer tool · Shipped for personal use
Clean-room pipeline
A clean-room rebuild pipeline with a gate that blocks decompiled code from crossing into the spec.
- Problem
- Clean-room rebuilds rely on people promising not to copy. Nothing enforces the boundary.
- My role
- Builder: the leak gate, the parity reporter and the analysis setup, including building the Mac decompiler from source.
- Shipped
- A CLI pipeline, a worked example and a reusable agent skill.
- Stack
- Python
- Ghidra
- Java
- MCP
- Claude Code CLI
- Outcome
- 28 gate tests
- Worked example rebuilt from a gated spec and matched the original on 50 of 50 behavior checks
-
Livestudio5ny.co
Studio 5 NY Custom redesign for a New York production companystudio5ny.co
Client website · Live
Studio 5 NY
A custom redesign for a New York production studio, built on top of Squarespace so the client keeps their editor and store.
- Problem
- The studio wanted a look of its own without leaving Squarespace or its store.
- My role
- Designer and front-end developer for Elevare Solutions.
- Shipped
- A live redesign on the client's own site.
- Stack
- HTML
- CSS
- JavaScript
- Squarespace code injection
- Outcome
- Live since October 2026
- Beat products stay hidden until the client confirms prices, then switch on
-
Livepontebellahair.co
Ponte Bella Salon websitepontebellahair.co
Client website · Live
Ponte Bella
A salon website with a style gallery, reviews, product ordering and automatic color and treatment reminders.
- Problem
- A local salon needed a modern site and a simple way to bring clients back on schedule.
- My role
- Designer and developer for Elevare Solutions: site, forms, reminder functions and the installable web app.
- Shipped
- A live multi-page site with a reminder system.
- Stack
- HTML
- CSS
- JavaScript
- Serverless forms and scheduled functions
- PWA
- Outcome
- Live on the client's own domain
- Reminder emails go out on a schedule without anyone at the salon sending them
Featured build
Vigil, the system I build with
Running a client business alongside many builds meant no single place knew what was open, what was owed and what could safely be done without me. So I built one: a CLI, an always-on local server, a dashboard and native apps, plus a build department (architect, coder, tester, reviewer) that runs jobs in parallel git worktrees.
- Planan architect agent writes the job
- Codea coder agent builds it in its own git worktree
- Testa tester agent tries to break it; its tests are protected
- Reviewa reviewer agent reads the diff, then I decide
- 52 services behind one versioned HTTP contract
- About 1,300 Python and 270 Swift test functions, plus 28 eval scripts
- Agents draft and prepare; sending, posting, paying and deleting always wait for my approval
Building an AI product from the ground up?
I'm looking for a founding or early engineering role where I can own the product, the implementation, and the quality of what ships.
wbaez117@gmail.comLinkedInResume
Client projects: elevaresolutionspro@gmail.com. Open to engineering roles and new client work. Based in New York.