blog

things that surprised me, in writing.

• FSRS v6 in Reme: porting a memory model to a study app

  How my exam-prep app schedules reviews: the power-law forgetting curve, difficulty-stability state, the exact update rules, and two deliberate departures from the textbook algorithm.

  Aug 26, 2026

• Reading .apkg: importing Anki decks into a Flutter app

  An .apkg is a zip wrapping an SQLite database with HTML-laden fields. The anatomy, the parsing pipeline, the validation gates — and why imported cards start with fresh scheduling state.

  Aug 26, 2026

• Deterministic multi-agent simulation as a research method

  The reproducibility problem in social simulation is an implementation problem. A deterministic kernel turns it into an experimental one.

  Aug 21, 2026

• A personal AI stack on a 1.6GB VPS: operational numbers

  What a self-hosted agent, search engine, vector DB and ingest server actually consume — memory and money, measured.

  Aug 21, 2026

• Building the Kernel-Level Location Forger: A Design Blueprint

  the implementation blueprint behind the spoofing model — module, hook points, control channel, rust rationale, and the mvp path.

  Aug 20, 2026

• Spoofing at the Source: Kernel-Level GNSS Position Forgery on Android

  forging location where the data enters — a model of why kernel-level spoofing wins the GPS war and loses the physical one.

  Aug 19, 2026

• The encrypted envelope: HMAC-SHA256 and AES-256-GCM in a notification protocol

  A notification pipe where the server stores everything and understands nothing. Each layer, key, and constant, and the reasoning behind it.

  Aug 17, 2026

• What 13 days of phone notifications look like

  Measured properties of a real notification stream: 1873 events, two apps, 94% concentration in one, 14x daily variance, ~38% duplicates.

  Aug 17, 2026

• Working rules for coding agents

  The rules that actually changed agent session outcomes in my setup: analyze before search, interrogate unclear requirements, parallelize independent work, and close the loop with verification and recorded lessons.

  Aug 16, 2026

• Detecting watered-down AI API relays

  Five measurable dimensions for checking whether a relay actually serves the model it claims: capability, knowledge cutoff, tool discipline, cache honesty, and billing ratio.

  Aug 16, 2026

• Memento: an end-to-end encrypted notification pipeline

  Design notes on a phone-to-server notification collector: encryption choices, storage format, and the install problem that turned out harder than the server.

  Aug 15, 2026