I have owned digital marketing, run the operations from the ground up, and now architect AI automation workflows that do both.

A single-file installable app sitting over a live spreadsheet. It reads and writes through a secret-protected webhook, unlocks with a PIN sealed by AES-256-GCM in the Web Crypto API or a fingerprint through WebAuthn PRF, and schedules its own reminders on the device with quiet hours and a weekly cap.

• Installable, offline-capable, and it owns no server of its own
• Secrets sealed on the device, never stored in readable form
• Money map: in, out, net, with spending ranked by category
• A replication runbook, so a fresh instance can be handed to another owner
• Running cost approximately zero

Chrome only ever hands an extension the pixels currently on screen, so a full-page capture has to be assembled: scroll a viewport, screenshot, repeat, then stitch the stack onto one canvas and encode it. Built as an in-house replacement for GoFullPage, with no account, no upload, no watermark and no paid tier.

• Fixed headers and sticky bars appear once, at their natural place, instead of repeating down every screenful
• Lazy images are warmed by a full scroll pass before the first shot, and smooth scrolling is defeated so slices land where they belong
• Inner scroll containers are detected, so an app shell that scrolls a panel stitches that panel
• A PDF writer in about a hundred lines, no dependency: PDF's /DCTDecode filter is JPEG, so each page embeds a JPEG byte for byte
• Only activeTab, so it has no standing access to any site and none at all until the button is clicked
• 38 tests measure the output pixel by pixel against 42 known color bands, which is what proves the slices landed rather than that a tall image came out

An Android progressive web app records in-person meetings and syncs privately to a laptop watcher, which transcribes with faster-whisper large-v3 on the GPU and returns a structured note to the phone. Audio and transcripts stay on the owner's own devices by design.

• Records on the phone, transcribes on a GPU at home, nothing sent to a cloud service
• Mic level meter, quiet-voice boost and noise cut before anything is captured
• Notes land back on the phone, and optionally into a chosen folder or Discord
• Also accepts voice memos recorded on a watch

A recorder captures screen, OCR and call audio through an expert's shift, and an analyzer turns the session into an explicit playbook a dispatch engine can implement. Local-only by default, with a separate and deliberately scrubbed cloud opt-in.

• Runs hidden from a desktop shortcut, so it does not sit in the way of the work
• Screen, text and speech captured together, which is what makes the reasoning legible
• Output is a written playbook, not a model: a human can read and correct it
• Nothing leaves the machine unless the cloud flag is passed explicitly

A Cloudflare Worker and a D1 queue that dispatch headless AI build tasks to a laptop runner from a phone, with token-checked actions, per-project routing, local-only logs and status notifications. The plan-and-execute pattern, built as infrastructure.

• PIN-gated, and every action is token-checked before the runner will accept it
• The queue holds the task list; the laptop collects work within two minutes
• Per-project routing, so a task can only run in the folder it was scoped to
• Logs stay on the machine that did the work