Built at night, while Oscar sleeps.

I'm Peter. I build small iPhone, iPad and Mac apps — for photos, storage, notes and wedding planning — and I ship them on my own.

On-device when the data is yours, relayed through a server when it isn't.

See the apps ↓ peter@calmdownoscar.com ↗

App Store

Wedding Concierge

AI wedding-planning concierge with grounded vendor search and shared state.

macOS iOS

Mac App Store

Unli Rice

Persistent memory layer for multi-agent LLM workflows connected via MCP.

macOS iOS

App Store

Unli Rice Capture

One-tap voice capture with on-device transcription that syncs to macOS.

iOS

App Store

ClearSpace

Native Swift/PhotosKit organizer with duplicate and blur detection.

iOS

Mac App Store

Unli Disk

macOS storage map and dev-cache analyzer that sweeps libraries safely.

macOS

Web

OpenGrail

Visual atlas of comparative religion and mythology across 695 traditions.

Web

Held for Polish

Shuttle Vision

Turns iPhones into an offline badminton scoreboard with slow-motion VAR replay.

iOS iPadOS

Awaiting Review

Kitchen Vision

Pickleball scoring and VAR review with automated 3-number callouts.

iOS iPadOS

Apps & Case Studies

How each app actually works: what runs on-device, what talks to a server, and why it's split that way.

Live on the App Store — iPhone, iPad & Mac

Wedding Concierge

macOS iOS Vercel Relay CloudKit

An iOS/macOS wedding-planning concierge — not a bare chatbot. The LLM is one piece of a system that also handles sync, credentials, and grounding.

Wedding Concierge: the AI Concierge on iPad, answering a bus-rental question with real local operators, capacities and prices rather than invented ones

Wedding Concierge on iPhone: the wedding board, with the venue and photographer already booked and the guest list, seating and budget still open

01 · Grounding & state

Grounded, not generative

Venue and vendor picks come from Apple MapKit and Google Places, so the assistant reasons over real search results instead of inventing addresses. The model and the UI also share one state model by design — a known failure mode in AI apps is the assistant's understanding drifting from what's on screen, and that's specifically what this was built to prevent.

02 · Relay architecture

No key in the binary

Cloud calls go through a stateless relay hosted on Vercel, which holds the API key and enforces rate limits server-side — nothing in the client can leak a credential. An earlier build ran inference on-device via MLX (Llama 3.2 3B, then Qwen 2.5 3B); that path was later pulled, and the app currently runs a single hardcoded cloud preset (Gemini Flash) with no on-device fallback.

03 · Status

Local-first, minimal data

SwiftData plus CloudKit keeps plans in sync across iOS and macOS without a backend database. No user tracking; sensitive profile data stays on-device. Live on iPhone, iPad, and Mac as of August 7, 2026.

nuptia.wedding →

Mac App Store · Open Source

Unli Rice

macOS iOS MCP Protocol Multi-LLM State Open Source

A persistent memory layer for multi-agent LLM workflows, built because every new AI session starting cold was burning tokens and causing hallucination loops.

Unli Rice on macOS: the Needs You screen, where the tidy-up helper surfaces things needing your OK — it can label and flag duplicates but never merge, change, or remove anything without approval

01 · Why

The cold-start problem

Every new session with Gemini, Claude, Codex, or Antigravity starts with zero context, so it re-explains the project from scratch. That's wasted tokens — and worse, a source of hallucination when the re-explanation is incomplete.

02 · Architecture

One vault, several agents

Unli Rice connects those LLMs over Model Context Protocol (MCP) to a single ground-truth vault. It's not just shared notes — the vault structure encodes explicit guardrails and evaluation rules, so one agent can't step on another's work or violate a constraint a different agent already set.

03 · Reach it

Voice in, code out

An iOS companion app captures raw audio and syncs it to macOS for retrieval by any connected agent. Source and guardrail design are public on GitHub, linked below.

Product Page → Source on GitHub →

Live on the App Store

Unli Rice Capture

iOS On-Device Transcription Free & Open Source

Built because I rarely have two free hands — my son follows me around like a shadow, and typing an idea on a phone one-handed usually means losing it. So I talk it instead: one tap, transcribed on-device, and it’s done. No rewriting it later — the note is already there on my MacBook and readable by whichever LLM I want to think it through with.

Unli Rice Capture: the welcome screen, explaining one-tap recording and on-device transcription

Unli Rice Capture: captured thoughts already listed as synced, readable notes

Product Page → Source on GitHub →

Live in App Store

ClearSpace

iOS PhotosKit Swift / SwiftUI 100% Local ML No Telemetry

Native Swift/PhotosKit rebuild, born from my own 40,000-photo library — real, not hypothetical. Now live on the App Store.

ClearSpace: the full timeline sorted biggest first, each thumbnail tagged with its size, and 4,535 unhearted items totalling 22 GB

ClearSpace: batch selection, with keep, heart, add-to-album and delete offered as separate actions

ClearSpace: the core loop spelled out — sorted by size or month, heart to keep, then clear the rest

01 · Detection method

Signal, not a model

What ships today isn't a model call at all: a 64-bit perceptual hash surfaces candidate duplicate pairs, Apple's Vision feature-print distance confirms the match, and a Laplacian blur score catches soft shots — all through Apple's Vision framework, no model download. An earlier version added an on-device VLM as a refinement pass on top of that; it made the scan slower for little gain, so it was cut. Classical signal was the better call.

02 · Safety

Non-destructive by design

The core flow is blunt on purpose: pick a period, heart the photos to keep, then one button deletes everything in that period without a heart. Skip the period filter and that button can queue 30,000+ photos at once — so every deletion routes through iOS's own Recently Deleted instead of a hard delete, giving up to a 30-day window in Recently Deleted to undo that mistake.

03 · Cross-platform Pro

Mac companion sync

ClearSpace 1.2 is live with previews that open from a prebuilt thumbnail cache, pinch-to-zoom inspection, and finishable monthly cohorts. If you own Unli Disk for Mac, ClearSpace Pro is automatically unlocked on your iPhone via iCloud.

Product Page → Privacy Policy

Live on the Mac App Store

Unli Disk

macOS AppKit / SwiftUI PhotosKit Core Dev Cache Engine

Native macOS storage map and dev-cache analyzer, built to find what generic visualizers miss and sweep photo libraries safely. Includes ClearSpace Pro for iPhone.

Unli Disk: the Disk Map view, showing reclaimable space alongside a measured directory breakdown

01 · What it finds

Not just folder sizes

Storage tools that only show what's big don't know a build cache from an uncommitted repo. Unli Disk parses specifically for orphaned Git repos and dev caches that naive visualizers surface as generic "large folders."

02 · Safety rule

Trash only, never unlink

Automated cleaners are a real data-loss risk on a machine full of uncommitted work, so every deletion routes through macOS Trash — recoverable, never a direct delete. That constraint was chosen on purpose, not left out for later.

03 · Includes ClearSpace Pro

Shared engine, cross-platform Pro

The photo-auditing code lives in Packages/UnliDiskPhotos, the same engine ClearSpace uses. Purchasing Unli Disk ($14.99 one-time) automatically unlocks ClearSpace Pro on your iPhone via iCloud.

Unli Disk Page → Privacy Policy

Live on the Web · Open Source

OpenGrail

Web React / Vite Static — No Backend Canvas Force Graph MIT

An open visual atlas of comparative religion, theology and ancient mythology — 695 traditions and 1,088 relationships you can actually move through, instead of 695 encyclopedia pages that never link to one another.

OpenGrail: the force-directed atlas with all eleven clusters laid out and the major traditions labelled

01 · Source of truth

Markdown in, graph out

Every tradition is a Markdown file with validated frontmatter. A build script resolves the links between them, computes backlinks, and deterministically regenerates a single graph.json — so the interface never needs a content API, and a correction is a pull request against a text file rather than a write to a database nobody else can see.

02 · Sourcing rule

Coordinates are resolved, never remembered

Each origin pin is two separable steps: an editorial decision about which place a tradition came from, and a mechanical lookup of that place against Wikidata, recording the entity ID that supplied the numbers. That second step is not ceremony — writing coordinates from memory had produced a Petra that resolved to a particle accelerator and a Tiberias in Indonesia, and the review catches that class of error before it ships.

03 · What it costs to run

Nothing to trust

The atlas is a static bundle: no backend, no analytics, no accounts, and no AI service in the loop when you read it. Epistemic-tier filters mark how well attested a claim is instead of flattening scholarship and folklore into one confident voice. MIT licensed, with the full record set public.

Open the Atlas → Source on GitHub →

Approved · Held for Polish

Shuttle Vision

iOS / iPadOS 100% Offline P2P VAR ¼× Slow-Mo BWF Scoreboard 8-Court Desk

Built because nobody in a casual game can reliably say whose serve it is, which box it comes from, or what the score actually is. Turns the iPhones you already own into a scoreboard that keeps the rules for you, with slow-motion replay for the rallies that are genuinely disputed — 100% offline.

Apple has already approved this build — it's sitting ready to go. I'm holding the release myself for a bit more tuning, because I'd rather ship it right than ship it on schedule. A new build is coming soon.

Shuttle Vision: the match scoreboard, showing which team serves and from which box

Shuttle Vision: the tournament desk on iPad, running live courts, fixtures and a standings leaderboard

01 · The problem

Nobody agrees on the score

Badminton's scoring is easy to state and hard to hold in your head. Serve passes to the other side on every rally won, the server's box flips between right and left with their own score, and doubles adds a pair rotation on top. A few rallies in, nobody is certain who serves, from which box, whether it's 11 yet, or when to change ends — so play stops while four people rebuild the score from memory. Whether the shuttle landed in or out is rarely the argument; badminton lines are far easier to call than tennis.

02 · Keeping the rules

Serve, box, ends and score

Tracks server, receiver and the diagonal service box across singles and doubles, prompts the 11-point interval and the change of ends, and handles official BWF Law 14 LET. When a rally is genuinely disputed, Review opens the last one from rolling memory in ¼× slow motion with frame stepping and 1-tap +1 / LET — video for the times someone's memory is conveniently wrong.

03 · Rig & desk

1 to 3 phones, up to 8 courts

Pairs two net-post cameras with an optional umpire-chair camera for sub-millisecond clock sync and Hawkeye 3D ray tracing, peer-to-peer with no Wi-Fi router. For club events, court umpires scan a QR code to stream live scores to a central iPad that computes official BWF tie-break standings in real time.

Shuttle Vision Page →

Awaiting Review

Kitchen Vision

iOS / iPadOS Side-Out & MLP Rally 0-0-2 Callouts VAR ¼× Slow-Mo 100% Offline P2P

Engineered specifically for Pickleball. Solves the doubles scoring puzzle by automatically computing Server 1 / Server 2 rotations, court parity, and 3-number callouts ("0-0-2"), paired with instant slow-motion rolling video review and multi-court tournament master desks.

Kitchen Vision: the pickleball scoreboard at the opening serve

Kitchen Vision: the casual hub, offering a quick match or living-room practice

Kitchen Vision: the tournament dashboard tracking two live courts at once

01 · The 0-0-2 Puzzle

Resolving the 3-number callout

Pickleball doubles scoring has a unique opening serve rule: only one partner serves at match start (0-0-2). On side-outs, both players serve in turn (Server 1 then Server 2), swapping left and right boxes upon winning points. Kitchen Vision tracks the exact serve number, receiving positions, and audible callout automatically.

02 · Handheld Umpire & VAR

Instant ¼× slow motion on disputed dinks

Designed for fast one-thumb scoring during matches. If a kitchen-line violation or close sideline call is disputed, tap Review to inspect the rally in ¼× slow motion directly from rolling memory with millisecond frame stepping.

03 · Mesh Rig & Master Desk

Multi-camera sync & live tournament standings

Synchronize 2 to 3 iPhones along the net posts via offline P2P mesh for multi-angle VAR review. Club organizers can run entire multi-court round robins and elimination brackets with live QR-connected umpire scoring.

Kitchen Vision Page → Privacy Policy

The multi-agent workflow →

About

I'm Peter. Oscar is my son — he's the reason this studio has a name, and he's still working on the calm part. The apps that make it to the App Store are built out of real necessity: tools born to solve my own development friction, or engineered to balance everyday life while raising a family.

My background spans enterprise systems, systems architecture, and wiring LLMs into relays, on-device models, and structured data so they behave predictably. When not shipping software, I'm usually thinking about quiet roads and riding motorcycles.

Apple Developer → GitHub → peter@calmdownoscar.com → More about me →