Diego Agustin Perez

Technical Program ManagerAI Builder

I ship real software by directing AI.

By day I run technical programs that serve millions of users globally. On my own time I run a small studio of AI agents: six apps that run parts of my life, the design system and brand practice behind them, and the process that keeps all of it honest. This is the build log.

Apps built and running
6
Commits
1233
Claude Code skills I wrote
13
Agent sessions
156
Mothership: Focus: the one thing first, then what's waiting and what's ready
Mothership: A project: its steps, its plan and its architecture
Macro Meeter: History: weight, then intake against each day's target
The Arrs: Postarr: every title with its poster, one wall
The Arrs: Reconcilarr: a finding, every system's view of the title, and its fix
Cue: On sale, film first: on sale, coming soon, not available
Pocket Change: Home: net worth, the biggest movers and runway
Pocket Change: Trends over 1, 3 or all years, from $0
Mothership: Every project, grouped by status
Macro Meeter: The Tracker: a whole day, the ring and four meals
Macro Meeter: Add to Lunch, frequently used first
The Arrs: Postarr's Inbox: new titles, each with its candidate posters
Cue: Alerts, grouped by day, each with its live status
Cue: The film screen: Watch At switches per house, showtimes, Your Dates
Pocket Change: This Month: last month beside each field, and the change as you type

Selected Work

Built for my own life, held to a professional standard.

Every app solves a problem I actually have, runs in production on my home server, and was built end to end with Claude Code. None of them are demos.

Mothership: One screen for everything I build and run: what needs me next across every project and machine, ranked, with a Start button that opens the next step as a Claude Code session on my Mac.
Focus as an orbit: your call, ready, quiet

01Mission control for my AI work

One screen for everything I build and run: what needs me next across every project and machine, ranked, with a Start button that opens the next step as a Claude Code session on my Mac.

  • The server never tells the Mac what to run. A launch names a project and a step id, nothing more.
  • A contract, not a convention. The status file has a schema, examples and a stdlib-only validator, so any repo's session can check its own file before committing.
  • Nothing is a stopgap. The Mac agent reaches the server through Cloudflare Access with its own service token and policy, not a LAN shortcut that would break the day I'm away from home.
In daily useFastAPISQLiteReact 19
Read the case study →
Macro Meeter: A macro and calorie tracker where logging a meal is a photo or a sentence: a model proposes each item with a portion and its macros, I check and correct it, and it lands in the day.
Check the Estimate: each item, how sure it is, and a question

02A food tracker that reads the plate

A macro and calorie tracker where logging a meal is a photo or a sentence: a model proposes each item with a portion and its macros, I check and correct it, and it lands in the day.

  • An estimate writes nothing. The estimate endpoint only proposes.
  • Targets with effective dates. Changing a goal never rewrites history: each day is graded against the target that applied that day.
  • Honest averages. A half-logged day can be marked Left Out.
In daily useFastAPISQLiteReact 19
Read the case study →
The Arrs: Postarr chooses a poster for every film and show and puts it on Plex with one click and an undo. Reconcilarr checks that five systems agree about every title, and turns each disagreement into a fix you preview first.
Reconcilarr's Inbox: only what needs me

03Twin apps for a 3,500-title Plex library

Postarr chooses a poster for every film and show and puts it on Plex with one click and an undo. Reconcilarr checks that five systems agree about every title, and turns each disagreement into a fix you preview first.

  • "A dead source is never a clean result". Collectors raise an error rather than return a short list, and anything under a sanity floor counts as down.
  • "Never assert an authority you do not have". Plex is wrong about IDs often enough that a conflict is shown as a choice between two resolved titles ("64691: a 1946 nutrition film" vs "64690: Drive, 2011"), not an automatic correction.
  • Matching that verifies, never guesses. The old matcher picked the wrong page for 3 of 77 sampled films and flagged none.
In daily use · v2 liveFastAPISQLiteReact 19
Read the case study →
Alerts: every push, grouped by day, with live status

04On-sale alerts for premium screens

A self-hosted watcher that pushes a notification to my phone the moment IMAX 70mm or Dolby Cinema tickets go on sale at the three LA theaters I go to, with one tap to the seat map.

  • "Getting flagged kills the project". Jittered intervals, conditional GETs, Retry-After honored, an hourly budget with a hard cap, a circuit breaker, and backoff persisted to disk so a reboot can't cause a hammering.
  • A trade-off, written down. Fandango's robots.txt disallows the endpoint Cue reads.
  • At-most-once alerting. After a flapping showtime fired about 1,000 pushes in a day, alerts moved to a ledger committed only after a successful send.
In production · formerly imaxerPythonhttpxPlaywright
Read the case study →
Pocket Change: A net-worth tracker for a household: twenty balances logged once a month as a checklist, trends over years, and a one-page report written for someone who isn't a numbers person.
This Month: a checklist, grouped by where you log in

05Household money, once a month

A net-worth tracker for a household: twenty balances logged once a month as a checklist, trends over years, and a one-page report written for someone who isn't a numbers person.

  • Encrypted at rest, LAN only. The database is SQLCipher-encrypted, with the key on the server's boot drive and never in the repo.
  • Backups that are proven. Encrypted copies every night on the server, pulled daily to the Mac, and a restore actually run and checked, not assumed.
  • Migrations proven on a copy. Schema changes are additive, guarded so they can run twice, and checked with a script that compares net worth month by month before and after, on a copy of the real data, before merge.
In daily use · LAN onlyFastAPISQLCipherReact 19
Read the case study →

The Build Log

1233 commits. Around the clock.

Every commit across the repos shown here, by project and day: 143 on the busiest day alone. 65 of them landed between midnight and 6 a.m., while agents I'd briefed before bed kept working. Hover a day for its count.

Last 6 weeks

The ArrsPostarr · Reconcilarr · shared
617
Mothershipmission control
108
Macro Meetertracker + meal estimates
113
Skillsmy Claude Code skills
115
Cueticket alerts · was imaxer
161
Pocket Changehousehold money
119

The Method

The code is the easy part. The rigor is the work.

I've written thirteen Claude Code skills that encode how I want software designed, audited and reviewed. Every change runs the same loop.

  1. 01Brief
  2. 02Audit
  3. 03Mock up
  4. 04Build in lanes
  5. 05Review
  6. 06Ship and log
The full method →

Why A TPM

Agents need what every program needs: scope, gates, and someone who owns the outcome.

Scope and requirements

Briefs with non-negotiables written before the work: "A dead source is never a clean result." "Never assert an authority you do not have."

Risk management

Live data is off-limits by rule, so tests run on copies and throwaway data dirs. Backups come before migrations. Anything that costs money asks first.

Decision records

Every brand round and design call is logged with its reason, so a later session can't quietly undo it.

Resourcing and budget

Model and effort chosen per phase. Cheap models do mechanical work in parallel, and usage is measured per model to estimate the next plan.

Parallel workstreams

Big tasks are split into lanes that run at once in subagents and git worktrees, then merged and reviewed.

Quality gates

"Done" means screenshotted at two widths and reviewed against a checklist, plus a construction review for any logo. Tests run before every commit, and a failing one stops a deploy.