Expanding what human ingenuity can do.

Obsidian · MemPalace · Markdown

I built a personal memory system with Obsidian for linked Markdown notes and MemPalace for retrieving saved knowledge across AI sessions. The notes remain readable and editable outside the retrieval tools.

What I learnedPersistence makes earlier work available, but it also preserves assumptions that may no longer apply. Keeping the notes in Markdown gives me a record I can inspect and revise when a decision changes. I need to distinguish a lasting fact from a temporary plan and retain enough context to understand why either was recorded. That review is part of maintaining the memory, even when retrieval itself works well.

Markdown · Semantic search · Source checking

I built a personal tool for searching Markdown research notes by meaning and tracing results to their sources. The returned material still needed review for relevance and context.

What I learnedSemantic search helps me find relevant material when I do not remember its exact wording. The result still needs to be read in context: a passage may discuss the right topic without supporting the particular claim I am making. Tracing an answer back to its source lets me check its scope and qualifications. I treat retrieval as a way to locate evidence, then make a separate judgment about what that evidence supports.

Whisper.cpp · FFmpeg · Subtitle timing

I used Whisper.cpp and FFmpeg in local transcription, translation and subtitle experiments. I reviewed uncertain wording and checked the timing and readability of rendered captions.

What I learnedEach stage introduces a different kind of error. A plausible transcript may contain the wrong word; a fluent translation may shift the meaning; accurate captions may still be difficult to read at the pace of the video. I need to check uncertain wording against the recording and judge the finished subtitles on screen. Running the tools locally changes where the processing happens, but it does not remove those editorial checks.

Codex · Python · Mac/Linux

I built a personal workflow that sent a self-contained task from my Mac to a Linux machine running Codex, then brought back the resulting files. I reproduced the worker’s tests on the Mac and checked the output independently.

What I learnedA separate AI session needs a task brief and inputs that make sense without the conversation on my Mac. Collecting files confirms that work came back; it does not establish that the task was completed correctly. Reproducing the worker’s tests and comparing the output independently gave me a stronger basis for accepting the result. I now think about how I will verify a delegated task when defining the task itself.

Python · Markdown skills · Dependency checks

My customized Skill Doctor tool checks reusable Codex instructions for missing files, unavailable dependencies and overlapping trigger descriptions. Its report identifies the affected skills and the issues to review.

What I learnedReusable instructions depend on files, tools and other instructions that can change independently. A checker can catch a missing resource or flag overlapping triggers before those problems appear in a task. It cannot decide whether the instructions are sensible or whether an overlap is intentional. The useful division is to automate checks with clear rules, then review the warnings and the meaning of the instructions myself.

Python · Unicode · DOCX validation

I used AI-assisted tooling to restore corrupted Sanskrit and transliteration in a Bhagavad Gita manuscript. The work rebuilt Devanagari and IAST text, preserved the edition’s 701-verse structure, and checked document links and rendered pages.

What I learnedChanging the font would not recover text stored with the wrong character encoding. Repair required comparison with reliable source text while preserving the edition’s wording, verse order and conventions. Structural checks could confirm links and verse coverage, but the rendered pages still needed inspection for readable characters and layout. Where the source remained ambiguous, documenting the uncertainty was more faithful than silently supplying a plausible correction.