What is running, what was tried, what is paused.

G2 reads "Go to". G2tiktokWork is the centrepiece of the family: a Telegram bot you give a niche and an idea, and which returns a finished 1080×1920 video. In between, a Python chain: generated script and hook, ElevenLabs speech with a local fallback, royalty-free background footage picked by keywords, Whisper transcription to align every subtitle word, beat-cut editing and FFmpeg rendering.

In Telegram the menu fits in four buttons: My Niches, New Niche, My Account, Quick Import. Each niche has its JSON configuration: voice, price range, its own footage. Users have accounts and credits in SQLite, backed up daily; the history of renders and used hooks avoids producing the same thing twice.

Actual state: in service, about sixty videos rendered. Publishing stays manual: that is G2TiktokAutopost's job.

The second member of the family takes the problem the other way round: the material already exists as a long video (YouTube, TikTok, Instagram). The bot downloads it, spots the segments that stand on their own, reframes them to 9:16 with a blurred background, burns Whisper subtitles in and returns the clips.

Two constraints shaped the architecture: heavy files, hence Pyrogram and the MTProto protocol to send up to two gigabytes, and a paid model, hence credits bought with Telegram Stars and a per-user processing queue. The same engine exists as a web dashboard, which served as a test bench before the bot version.

The last link before the viewer's screen. Rendered videos are filed by series, each series has its account, and the bot publishes at the chosen pace by driving a browser with Selenium over saved sessions. It dismisses pop-ups, respects rate limits and moves each published file into a posted/ folder so nothing is ever posted twice.

No interface beyond Telegram: that is where series status and errors arrive. It is the most fragile tool of the family, because it depends on an interface that is not its own.

The G2 family has one member turned toward live. A client listens to a TikTok Live chat, a game engine keeps questions, answers and scores, a WebSocket server pushes state to an overlay the streamer embeds in OBS, and a synthetic voice announces rounds.

Prototype: questions live in a JSON file, no accounts, no persistence beyond a session. It taught what a chat can do in real time, which fed Roast My Music.

Same base as the rest of the family, Telegram dashboard included, but a completely different ground. The bot listens to the Solana blockchain over WebSocket to spot new liquidity pools on Raydium, screens every candidate (honeypot, mint and freeze authorities, holder concentration), then executes, in simulation or for real, with post-trade rules: securing part of the gains, liquidating stagnant positions, dynamic stop and target.

It is shown here for the mechanics: real time, automated safety, asynchronous end to end. Not as advice, and it is paused.

The starting point is a live-stream format: a streamer listens to tracks or albums with their community and rates them on air. Roast My Music equips that moment. The streamer loads a track, gives a grade from 1 to 10, and the chat votes at the same time: Twitch and TikTok messages are read in real time, weighted (a TikTok gift counts more than a message), and the final score locks when votes close.

Three surfaces: the streamer dashboard (now playing, their grade, per-platform averages, live chat activity), an overlay to embed in the stream showing the moving score, and a history of sessions. A bridge listens to the chats through tmi.js for Twitch and a TikTok Live connector.

An older second track explored the classic album page with ratings and reviews. It served to test the rating flow, but the live format is what carries the product: a collective score in real time is what neither Letterboxd nor RateYourMusic does.

Actual state: working prototype locally, no accounts, sessions stored as JSON. Next step: a shared data model and Twitch authentication to open the tool to other streamers.

The idea was mine: the street is an open-air street art museum with no labels and no guide. Hall of Arts geolocates the works around you on a map, tells each one with text and audio, keeps a timeline of its evolution from community photos, lets artists claim and publish their works, and shows a work in augmented reality where it stands, even after it has been painted over.

We were three. I held the project lead role, from the brief to the delivery milestones, and the developer role on the React Native app: map and geolocation, work pages with an audio player, ratings and reviews, photo uploads. The first version ran on Firebase and Mapbox; the 2026 rework moves to a dedicated API and integrates the augmented-reality view with ViroReact. The project was presented as our final-year project and counted toward the degree.

A new version with a rebuilt back end was started in 2026. It is shown here for what it is: a completed school project, currently paused, whose augmented reality on heterogeneous phones remains the most expensive and most useful lesson.

A mobile prototype applying a simple method: seven days of observation to establish a baseline, then a staged reduction plan, with a daily log and the reason behind each intake. Charts show the gap between plan and reality.

It is a behavioural-health app in the same sense as a sleep diary or a smoking-cessation tracker. It extends, on another ground, the work done on Neocortex: giving a person and whoever supports them an honest measuring tool rather than a judgement.

The prototype never went beyond a local demo: no accounts, no sync, no clinical validation.