MC/
Start a conversation

Build log / Current work

Ideas are cheap.
Operating proof matters.

These projects show how I’m applying AI now. Status labels are intentionally conservative: a prototype is not production, a build is not a launch, and a passing test is not customer acceptance.

01

Browser intelligence

PageSift

A difficult input problem becomes a repeatable capture workflow with preserved structure, explicit permissions and recoverable archives.

  • Chrome MV3
  • TypeScript
  • On-device AI
  • Privacy
  • Release engineering
Private beta

Key decisions

  1. Rendered-page extraction before model involvement
  2. Schema-constrained Smart Lens with deterministic fallback
  3. Local-first capture with optional user-controlled destinations
02

Agent infrastructure

MCP Conductor

Makes multi-tool agent workflows easier to operate across Claude, Codex, Gemini and other MCP-capable clients.

  • MCP
  • TypeScript
  • Deno
  • Sandboxing
  • Agent operations
Public

Key decisions

  1. Isolated Deno execution boundary
  2. Cross-client configuration and recovery tooling
  3. Batching and result shaping to control context use
View public repository ↗
03

Native macOS systems

Lumnit

Turns fragmented display controls into one native surface while handling very different hardware channels safely.

  • Swift 6
  • SwiftUI
  • Hardware integration
  • Concurrency
  • macOS
In development

Key decisions

  1. Channel arbitration rather than one brightness mechanism
  2. Watchdog and stale-state recovery
  3. Separate direct and App Store capability boundaries
04

Human-controlled AI

Project Harmony

A working POC for authentication, family spaces, messaging, audit and export—with AI deliberately bounded behind safety decisions.

  • TypeScript
  • Next.js
  • PostgreSQL
  • Audit systems
  • AI safety
POC

Key decisions

  1. AI assists; people decide
  2. Hash-chained audit and explicit permissions
  3. Legal, safety and privacy uncertainty documented rather than hidden
05

Source-qualified automation

FlyOnTheWall

Creates queryable records while retaining immutable source snapshots, hashes, parser versions and unparsed rows.

  • Python
  • FastAPI
  • SQLite
  • Data provenance
  • Automation
In development

Key decisions

  1. Polite bounded collection—no control evasion
  2. Immutable raw evidence before normalisation
  3. Coverage gaps and source qualifications remain visible
06

Verified market tooling

IBKR MCP

Lets an AI workflow inspect market data without silently turning analysis access into trading authority.

  • Python
  • FastMCP
  • IBKR
  • Observability
  • Safety boundaries
In development

Key decisions

  1. Fail-closed analysis mode
  2. Qualification and provenance in every result
  3. Live integration evidence kept separate from unit-test evidence