Systems in various states of evolution.

The problem

Finding good deals by hand is fragmented, repetitive, and slow. Deal pages are noisy and inconsistent, and staying on top of them means searching the same places again and again.

What I built

A system that collects deals, processes the product information, rewrites the descriptions, organises everything into categories, and prepares the content automatically — an end-to-end pipeline rather than a one-off script.

Manual work it replaces

Instead of manually searching for deals, rewriting descriptions, sorting products, and assembling posts, the system takes over more of that pipeline with each iteration.

What I learned

This taught me what actually motivates me: building systems that produce value continuously, not one-time outputs. A single good post is forgettable. A pipeline that keeps producing them is the thing I care about.

Built with

Status

Still evolving — and honestly, the hard part now isn't building features. It's the unglamorous work of turning it into something fully operational and self-sustaining. Which is exactly the edge I know I need to push on.

The problem

Grocery pricing has almost no transparency. It's hard to compare prices over time, or to notice when something has meaningfully changed.

What I built

A tracker that collects and monitors pricing data — focused on historical tracking, structured data, and visibility into how prices actually move.

Manual work it replaces

Manual price-checking and the guesswork of “is this actually a good price?” — replaced with a record that just accumulates on its own.

What I learned

Even a fairly simple system becomes powerful once data is collected continuously over time. The value isn't in any one reading; it's in the history.

Built with

Status

Experimental / partially active.

The problem

Real discounted-grocery opportunities exist, but discovery and organisation are fragmented and scattered.

What I built

A set of experiments around deal discovery, structured aggregation, possible automation flows, and organising real-world discount data into something usable.

Manual work it replaces

The manual hunt-and-sort of finding and tracking discounts across a noisy, constantly changing source.

What I learned

Real-world systems are usually limited by operational complexity and data access far more than by the programming itself. The code is rarely the bottleneck.

Built with

Status

Exploratory / ongoing.

The problem

I didn't want to depend entirely on paid APIs and external platforms for AI work. I wanted local control, lower running cost, and the freedom to experiment without a meter running.

What I built

Local AI environments — ComfyUI, local LLM workflows, image-generation pipelines, and AI-assisted automation — aimed at generated content, image generation, workflow automation, and reducing dependency on outside services.

Manual work it replaces

External, pay-per-call AI services for the experimental, high-volume work where that model simply doesn't make sense.

What I learned

AI gets dramatically more interesting the moment it's wired into a real operational system instead of living in an isolated demo. The demo is the easy part.

Built with

Status

Highly active and evolving — the area I'm most engaged with right now.

The problem

Understanding how people actually talk about a topic — what they like, what frustrates them, what they'd change — means sifting through hundreds of Reddit threads by hand. No existing tool made it easy to pull structured sentiment from raw discussion at any scale.

What I built

A keyword analysis tool that scrapes Reddit's public JSON endpoints (no API key, no account required), collects every relevant thread and comment for a given search term, and produces a structured breakdown: pros, cons, suggested improvements, common complaints, and overall sentiment. A Streamlit web interface handles input, live progress, and download — all launched with a single double-click on Windows via a run.bat file.

Manual work it replaces

The manual process of opening dozens of Reddit threads, reading comment sections, and trying to synthesise a picture of what people actually think. The tool turns hours of reading into a structured, AI-ready output in minutes.

What I learned

Reddit's public .json endpoints are quietly one of the most underused data sources on the internet — no auth, no rate-limit headaches, and real human opinions rather than curated content. The heavier insight was about output design: structuring the data as AI-ready JSON and CSV from the start turned a standalone scraper into something that slots into any larger analysis workflow without modification.

Built with

Status

Built and working. The intentional simplicity — two libraries, no database, no API key — is the design decision I'm most satisfied with. It runs anywhere Python runs, which is the point.

The problem

Small repetitive tasks pile up into a real mental load over time — data entry, file handling, shuttling information between systems, the same operational actions on repeat.

What I built

Small automation scripts and workflows whose only job is to delete a repeated action from my day.

Manual work it replaces

The dozens of tiny manual actions that individually take seconds and collectively wear you down.

What I learned

Removing a few seconds from a repeated workflow gives back an amount of mental relief that's wildly out of proportion to the time saved. That trade is almost always worth it.

Built with

Status

Less a single project than an ongoing habit and a philosophy.