Things I wanted to use.

Why

I wanted to watch my bias through a whole group performance, even when there was no individual fancam.

How

• Use InsightFace face embeddings to recognise the selected performer from curated references, with confidence and movement checks.
• In enhanced mode, use YOLO11m person detection and BoT-SORT tracking to bridge short gaps, with face identity as the anchor.
• Interpolate and smooth the camera path to render a vertical crop with the original audio.

What

A vertical follow-cam video for one selected performer, with the original audio.

Ambiguous identities and long occlusions still need review.

Why

I wanted a repeatable workflow for translating videos, correcting subtitles, and keeping the segments worth sharing.

How

• Hermecho uses Whisper for transcription and OpenRouter models for translation with reference context, then writes timed SRT files and uses FFmpeg with libass for optional subtitle burn-in.
• Anamnesis wraps that pipeline in a local studio for subtitle editing, style presets, version history, and clip preview and export.
• AI-suggested ranges and manual clip selection help me keep useful segments in a searchable local archive.

What

Translated SRT files, subtitled videos, and selected clips kept in a searchable local archive.

Subtitle text, timing, and AI clip suggestions still need human review. The articles cover earlier iterations and experiments.

Why

I started with agents that helped me find and maintain my notes, then used skills to research new knowledge. Coding agents now let me try more ideas, and my need has shifted towards maintaining the content and projects those ideas create.

How

• Use Deep Agents with OpenRouter models for project check-ins and source-backed research through a Research skill.
• Read selected repositories through a read-only GitHub MCP server, keeping sources and open questions with suggestions.
• Save plans and decisions in SQLite, separate from chat history, and review suggestions before adopting them.

What

A personal PM agent in development, intended to help me keep content and side projects current as I explore new ideas.

Work in progress: the ticket board is a design preview. Live model and GitHub check-ins still need end-to-end validation.

Why

I am learning Korean and wanted a web app for my friend and me to practise the material from our class. I wanted our review to follow what we had actually been taught.

How

• Use React for the interface, with Cloudflare Workers and D1 saving course choices and review progress.
• Organise class material into vocabulary, grammar, knowledge cards, and practice questions.
• Choose what has been taught and what to add to review, with separate learning progress for each of us.
• Keep material sources and revisions visible so corrections can be reviewed before publication.

What

A private web app for class-based Korean study, with searchable knowledge cards, practice sessions, and saved review progress.

Course coverage depends on the material imported and reviewed. Explanations and practice questions still need checking.