kg Food Log

kg is a food tracking log that lets you use natural English language to log your food items.

It provides an LLM chat like interface to enter your food items.

I disliked scanning food items or finding items in a database, and taking photos of meals to allow AI to infer the items is not accurate enough. For example, olive oil is very dense and has a high impact on health, yet a photo will not convey that information.

I typically track my own foods in grams, and have found that just describing the item and adding “x grams” allows the LLM to accurately infer the nutrients in that serving.

LLMs are not currently completely accurate, but I believe in the upcoming years they will become more capable and faster, which would make this app more viable.

Currently it takes around 30 seconds to look up a new food item, which is not an issue as the user does not need to wait around. But almost instant responses would be an improvement.

It currently works best on desktops as it uses the full screen for canvas based visuals.

I have found canvas for visualizations much faster and capable of handling large amounts of data vs SVG or HTML visuals. Echarts and Glide Data Grid have turned out to be very good at handling large datasets.

I created a concept for sparklines drawn with bezier curves which came out much better than I had expected. These are drawn inside each Glide Data Grid cell, which allows for 120fps infinite scrolling without having to use the DOM API. I tested quite a few DOM based infinite scrolling APIs, but all produced jank. Glide Data Grid was exceptional here.

I find the curves of the bezier lines much easier to quickly scan to find highs and lows (vs reading every number). When using group by (food, meal, day, week), you can quickly scan for highs/lows of a given nutrient.

I find biochemistry fascinating, and with LLMs it is fast to research any areas of interest.

The goal of kg is not to impose any kind of dietary goal on the user. It is just a simple tool for looking up data, calculating sums and visualising them. You can set your own goals as needed.

I had originally chosen "Diet Snitch" for the name of this app.

You can find a demo with 6 months of food data loaded on the homepage.

To health 🥐

• kg.enzom.dev
• https://www.youtube.com/watch?v=a_6jQdxXRDo

Passkeybot

Passkeybot is a hosted sign in page that only supports passkeys.

It aims to allow you to add email verification and user auth to your app with only a few server side HTTP handlers.

Passkeys are quite a recent web API addition, becoming widely implemented in browsers since early 2024.

I believe the passkey APIs and native GUIs will continue to improve over the next year making creating and using passkeys much easier for end users.

I used the ROR (Related Origins Requests) feature for Passkeybot. This is a feature that allows domain owners to allow other domains to manage passkeys for their domain RP (Relying Party). The ROR feature is designed for one domain owner owning many related domains. But Passkeybot uses it more like a "passkey permissions API". Once a domain owner grants permissions to Passkeybot via ROR, passkeybot.com can create and use passkeys for their domain.

Passkey implementation is not particularly complicated, especially with LLMs. But the issue is if you have many domains, and you need email verification, you end up spending many days reimplementing, copying and pasting code, managing DNS, retesting the flows etc.

Passkeybot aims to make it a one-shot process where you add a few server side HTTP handlers and it gives you email verification, forgot password, passkey auth, and the sign in UI.

There is some amount of trust the domain owner needs in Passkeybot, but it can be reduced by re-verifying the passkey auth and email verification. In the future I would like to research using server side code attestation with an open source server implementation, so the domain owner knows exactly what server side code was running and will not need to re-verify their side. Golang seems like a particularly good runtime and language for this, as you can get bit-for-bit reproducible builds, the TLS and crypto libraries are very well tested and widely used, and the Golang compiler can create a single static binary without linking to C code which would make code attestation and review more straight forward.

There are also many upcoming Passkey APIs in development in the standard that are not currently released which I feel will make passkeys and Passkeybot more streamlined.

You can find more information here:

• https://passkeybot.com
• https://github.com/emadda/passkeybot

Hot Notes

• Hot Notes is a macOS app that allows you to open your Apple Notes from anywhere on your computer.
• It uses "fuzzy search".
  • Fuzzy search only enforces that all letters in the search term must exist in the correct order in any matches.
  • But: it allows for zero or more characters between the letters in the search term.
  • E.g:
    • Search: exar
    • Match: Extracting Artwork.
• Four letter searches. As a rule of thumb, when searching, I try to match the first two letters of two words in the title. This works quite well as the matching prefers matches at the start of word boundaries, and matches with a sequence of direct letter matches. So the title you are looking for will likely rank first.
• The reason for this app is that the Evernote native macOS app used to have something similar (cmd-j), but since the company was sold and moved to a purely web based solution, the performance got noticeably worse. Web based UIs can be designed to be fast (Figma, VSCode), so I do not think it is an issue with the technology, but rather the new company do not seem to value that "fuzzy search" feature.
• For me, the fuzzy search feature was the only way to actually find my notes in Evernote. I already have a vague idea of what I am searching for, because I originally wrote the titles, so I know roughly which words to search for. Also, I emulated "nested notebooks" in Evernote as it only allows one level of nesting (called "Stacks"). The way I emulated nesting was to name notebooks like so "Dev: Languages: JS: Misc". This worked exceptionally well with the Evernote fuzzy search as the entire path is included in the match as it is flattened into a single string. I will also experiment with something similar in Apple Notes.
• Nested note paths combined with titles are quite useful to fuzzy search for, because they have already been compressed by the user. Similar to how Twitter limits your characters, so you must reword it to be concise before you send it, note titles contain the essence of the note and force users to compress information before writing. This "forced compression" stage aids recall when it comes to searching for titles. Had someone else wrote your title (or an LLM), you would have no recall, and no idea what to search for. The note paths allow you to perform an implicit "branch chop" search where you exclude the non-matching branches, which excludes a large set of notes that are contained within them. Together (note paths and titles) searched via fuzzy search seems to be working well.
• Another reason for this app is that Apple seem to ignore folder searches in all of their UIs.
  • You cannot jump to a specific folder, or limit searches to notes in a specific folder.
  • Spotlight also ignores folders.
  • Folders cannot be referenced via ID in AppleScript, and must be referenced by name, which causes issues with nested notes. It forces you to traverse the nested folder structures to get a folder ref in code, when an lookup_folder_by_id("folder_id_abc"") API would work much better.
• I tried to use JXA, the "official" JS Apple Notes API, but it seems to be rate limited to one note every few seconds which was way too slow.
  • In the end the SQLite database was read directly, which other apps seem to do as well. This was much faster, nearly always less than 200ms.
• Existing apps.
  • There are many existing apps that search Apple Notes, but none seemed to work well for me. They had some combination of:
    • No folder search.
    • Requires Full Disk Access (which is not really needed, only the Apple Notes folder is needed).
    • Poor fuzzy matching.
      • E.g. A search with two words did not match a title with both words.
    • Very slow or non-functional with large numbers of notes.
    • Plugin systems.
      • The main app would download scripts from GitHub and run them, without ever strictly verifying the developers.
      • The Apple Notarization system is quite annoying to work with as a developer, but I feel the "hardened runtime" that avoids any dynamic code at runtime, and only allows signed code to run, works quite well.
      • In short: I prefer a signed app with minimal permissions than an app that can download and run any plugin code with Full Disk Access.
• Apple Shortcuts are used for the global shortcut, but surprisingly, this can be quite buggy. Some keyboard shortcuts cannot be registered for some reason, Shortcuts just ignores them. The "Open App" action can be slow at times. Raycast seems much faster in this respect. Apple Shortcuts was used to avoid having to give Hot Notes the keyboard monitor permission, again to limit permissions as much as possible.

Pocket Log

• Converts your audio logs to a visual timeline.
• Ideal for many short entries throughout the day, like a food or training log, or project notes you will return to.
• Allows copying your audio logs to paste into your LLM of choice for extracting, summarising and searching.
• Can transcribe Apple Voice Memos, and when synced to macOS via iCloud, allows taking memos via an iPhone, Watch or iPad and having them transcribed and added to the Pocket Log timeline.
• Runs in the background, so can watch for newly mounted USB drives like dictaphone or SD cards. When detected, it finds new audio files and adds those to the timeline.
• Saves data to an HTML file so the timeline can be viewed on any computer in the future without the app.
• Download Pocket Log on the App Store

Robata: put it on the grill

• This is a macOS app written in Swift and HTML/JS/CSS (Mobx/React/Tailwind).

• I created it because the default "Mission Control" window manager for macOS has some issues:

  • A. It is not possible to quickly jump from one space to another when there are other spaces between them.
    • The "three-finger side swipe" trackpad gesture only works well when swiping to a neighboring space, and when you know what you will find on the space.
      • Otherwise, you end up paying the cost of each animation as you swipe between each space searching for the window you are looking for.
  • B. "Expose" does not work well when you have more than 3 windows on a space.
    • This is the "three-finger swipe up" trackpad gesture.
    • The animation seems to "explode" the windows to show them all, but this is quite distracting and unpredictable. The more windows you have, the smaller each one will be in this view. The animation also takes a few seconds to complete.
    • The fact that windows are semi-randomly placed and small makes it difficult to visually scan them.
  • C. The top "space viewer" is so small you can barely tell which windows are on those spaces.
    • It also requires that you jump through two animations to get there "three-finger swipe up", and then mouse over the spaces row at the top of a monitor.
  • D. Stage Manager does not seem to work well with spaces, adding another layer of groupings and UI when a better interface to the existing spaces could work well.
  • E. The default cmd-tab of macOS switches between apps and not windows. It also orders by most recent, but often the window I want to switch to is around 4-6th place depending on the tasks I am currently working on.

• The default macOS animations for spaces are nice, but when they take a few seconds and are unpredictable, I think they have a negative effect on the user experience. A good principle for animations seems to be that if they are predictable, they are good, but if the user has no idea what the animation will reveal, it makes for a worse experience as the user has to wait for the animation, then spend a lot of visual processing time working out what information the animation is presenting. You cannot disable the animations for spaces on macOS.

• Robata has minimal animations and shows windows instantly to avoid the double-edged sword of animations. Humans have very fast visual pattern matching, which is why I want to show the actual windows, as the patterns of the window UIs for different apps should be instantly recognizable.

• I also do not like the icon-centric design of the default macOS cmd-tab. I feel scanning for the right icon is a waste of time, as you then have to scan for the right window contents to select the window you are looking for. It is better just to remove the icons and keep the window screenshot and title, as these are more specific to the task you are actually working on in those windows. An example is Chrome—you might have 10 tabs open, so the icon does not give you any clue as to which tab you want.

• The app uses private macOS APIs to get data and events on spaces and raw trackpad events. These are undocumented, so they require a lot of experimenting with. I feel Apple should support and create documentation for these APIs so users can develop UIs to use their own computers how they like.

• You can activate the Robata UI by resting a finger on the corner of the trackpad. My thumb naturally rests on the bottom left corner of my external trackpad, so it's very natural to rest it there to open Robata. Also, the tactile feel of the corner of the trackpad makes it easy to find without thinking. An issue with the MBP built-in trackpad is that it is flush with the body of the laptop, so it takes more effort to find the corner. So, I added the ability to use cmd-{x}—where x is any char the user picks. This lets you open Robata with just your right hand on one half of the keyboard, or just your left hand on your trackpad. The key shortcut is also for mouse users.

• The Robata app is around 1.8MB and has no dependencies (other than the macOS system libraries). It uses web views, which turned out to be very performant. As soon as the web views are no longer visible, they go into a sleep mode, reducing their RAM and CPU usage. Robata is not visible for most of the time, so this works well.

• Some Swift developers have noted that text rendering using SwiftUI actually takes a lot of CPU when the text changes every second. I feel the optimizations built into web browsers make for a more performant UI in some cases. Also, the instant reload and Chrome dev tools make it much easier to develop with. You can change your text and see the re-render instantly, like painting a canvas. With SwiftUI, it is a minimum of a few seconds to recompile and navigate to the app state you are working on. Also, SwiftUI layout seems to be harder to debug than HTML/CSS, as the Chrome dev tools will show you why your elements are positioned where they are with flex/grid ruler overlays. SwiftUI has no debug tooling.

cssprinciples.com

• I built this blog to teach CSS principles.
• The CSS grid guide is probably the most interesting post to date: cssprinciples.com/3/
• Both the static site and the examples are built with React/MobX.
• The static site generator is a custom built script using Bun.js.
  • I chose Bun.js for its performance, it's particularly fast at hot reloading to preview changes instantly in the browser after an edit.
• The content is written directly in JSX to maintain full control over HTML output, which is needed to allow maximum interactivity and customization in the future.
  • I avoided Markdown and MDX (JSX in Markdown) because:
    • The MDX compiler has less users than JSX; I think MDX based builds will be more likely to break in the future.
    • It seems more difficult to customize than just using JSX directly.

tabserve.dev

• Gives developers an https url for their localhost webserver.
• Enables them to share their local web server with the world, receive API webhooks, test on different devices and use https-only browser APIs.
• More detail at tabserve.dev

bigWav

• This is a privacy first transcription app.
• It allows you to annotate your transcriptions and export them to PDF.
• You can also save as .bigwav.

Technical details

• Uses OpenAI whisper to convert audio to a transcript.
• Whisper.cpp allows you to run the processing in the browser via WASM.
• Uses Chromes File System API to read and write to files on the file system.
  • This is useful as the app has its own file format (.bigwav) which stores the audio along side the transcript and annotation.
  • When the user saves the file, it is written to their disk similar to how other local apps work.
• Is an installable PWA.
  • This is useful for opening audio from the users file explorer.

Initial Instinct

You can learn more about the rules of the game at initial-instinct.com, tap "about".

Some technical details about the game:

• Extracts peoples names from Wikipedia metadata and page content using Rust, inserting them into a SQLite database. The Wikipedia database is a single large XML file, so using Rust allows for faster iteration/changes as you can inspect the export results 10x faster.
• Uses Node.js on the server, with SQLite, Google Container OS, Caddy for auto HTTPS.
• Frontend uses React with Mobx.
  • The animations are videos as they were designed in Adobe After effects. To get a 1:1 replication it was easier to use the video's of the animations as 3d is hard to replicate using CSS animations.
  • This took a bit of effort to get running smoothly on iOS/iPhones as they try to prevent any caching to save battery, RAM and CPU. The wheel animations need to start at the exact correct moment, and iOS tries to lazy load them at the last moment, which can cause a multi second delay.

Stay 100

Stay 100 is a web app that shows you 100% of the stays for a given date range and location.

It is optimized for large screens by using all of the available space to show images.

At the time of launch, AirBnB paginates results making it difficult to quickly visually scan for interesting stays. This can mean missing stays because you did not look through all of the result pages. AirBnB also does not allow sorting or filtering by all fields (like rating, number of reviews or price).

Explained in more detail on https://stay100.madebyenzo.com.

Table Dog

• See https://tabledog.madebyenzo.com for details.

Note To Web

• Allows you to create a website from an Evernote notebook.
• Includes a category system based on naming the notebooks with a tree path.
• Static hosting with a CDN is used for extremely low latency serving of web pages.
• Custom domains supported.

SQLite MPI

As part of Simple Job Tracker I needed to use SQLite, but the open source libraries available had dubious support for transactions and cross platform reliability.

In essence it uses Rust to cross compile a single code base, instead of re-implementing the same code in both Java and Objective C.

It turns SQLites FFI into a JSON RPC so that it can be used over the React Native bridge (the interface between the JS VM process and the native app process).

SMPI is described in much more detail at sqlite-mpi.madebyenzo.com.

Simple Job Tracker

SJT is a mobile app for Android and iOS that is intended to be used by small business owners.

Its designed for location based recurring service businesses. This means any business that visits many addresses per day, and will visit those same addresses on a recurring schedule in the future.

It manages invoice and receipt numbers, customer balances, customer contact details, upcoming jobs.

Data is synced to a remote store, allowing many users to use the app in a single business.

The SQLite MPI was designed as a result of this project.

SJT is described in much more detail at simplejobtracker.madebyenzo.com.

Tree merge

Tree merge is a tool I built in order to merge the category systems for all the DIY stores in the UK (see Everydiy).

It increased conversion rates by 100%.

The original issue was that when searching with vague terms, Elastic Search would return too many products.

Using search plus categories allows the user to very quickly drill down to what they are looking for.

Tree merge is described in much more detail at treemerge.madebyenzo.com.

Everydiy

Note:

• This project was originally named "Lightcart".
• Treemerge is a tool that was built for this project.

What is Lightcart?

Its a search engine for DIY products available on diy.com (B&Q), homebase.co.uk, screwfix.com and wickes.co.uk. The idea is to enable you to shop at all stores from one fast and focused interface. Once you have a list of items you want to purchase in your cart you are redirected to the suppliers site’s checkout page ready for you to pay.

The target users are people who have many things to purchase and want a quick and easy way to see what is available at all the DIY stores without jumping from tab to tab.

The name “Light” was chosen after the 3 attributes I wanted the application to have:

• Fast
  • As in “lightening fast”.
• Comprehensive
  • Light enables sight.
  • Seeing results from all suppliers.
• Intuitive
  • As in “light entertainment”.
  • Easy to understand/use.

2010 research project revisited

This is the implementation of a research project I had for my degree dissertation in 2010 (titled “The possibilities and limitations of a meta shopping basket”). The original project was written in PHP with very little Javascript. Cross site XMLHttpRequests were sent with signed Java applets.

2015 technologies enabled a faster application

ReactJS/Javascript was used in a “single page” style application. ReactJS has a focus on fast DOM updates. The aim for the project was to keep application responses under 100ms.

A Chrome extension was used in place of a Java Applet to make the add to cart process much smoother.

Screenshots

transform-x

• https://github.com/emadda/transform-x
• JS functions to convert between SQLite, Excel and JSON.

Technical details

• This is a JS and WASM based library that works in any JS runtime.
• JS runtimes include: Node.js, Bun, Deno, browser, Cloudflare workers.
• Getting the same code base to work in all the JS runtimes was quite a challenge as they all have slightly different API's and have different security restrictions.
• Can be used as a CLI, JS library or as an HTTP API.

lilo

• https://github.com/emadda/lilo
• A CLI to download GCP logs to a SQLite db

Technical details

• Uses Bun, a new JS runtime written in Zig that replaces Node and uses JavaScriptCore, Apple's JS engine for Safari.
• This project was to test Bun out as a runtime.
• The SQLite writes seem to have good performance, taking around 10ms for 1000 rows inserted.
• SQLite is built into the Bun runtime, so there is no need for another build process.
• In Node.js, better-sqlite3 requires compiling SQLite which takes around 30 seconds and complicates distribution.

Durafetch

What Cloudflare Durable Objects are.

Cloudflare Workers allow you to upload JS code to respond to HTTP requests that run in 200+ globally distributed datacenters. This reduces user-to-server latency to less than 50ms regardless of user location.

Edge hosting solutions do not handle state well as there is no central location.

In typical web servers state is stored in a central location. This is the simplest possible set-up (single web server, single database), but has a few disadvantages:

• Will only have low latency for users close by. Latency can be 300ms for the opposite side of the world.
• Must be always-on to handle those requests (a small VM is around $12/month).
• Is a single point of failure.
• You must manage the OS updates, CPU/RAM/disk sizing etc yourself.

Durable Objects are defined in JS classes that are bundled with a Cloudflare Worker. The worker acts as a proxy to receive public HTTP requests and then instantiates a Durable Object to pass the HTTP request to.

Durable Object's have these features:

• A single identity
  • You can name a specific server instance with an ID and route requests to the same instance.
  • This allows you to keep JS state in memory.
• Persistent storage.
  • A JS key value datastore, like localStorage that is written to disk.
• WebSockets.
  • You can upgrade the HTTP request to a long-lived WebSocket for bidirectional communication.
  • This allows the WebSocket to interact with JS and KV stored state.

What Durafetch is.

Durafetch consists of two parts:

• A CLI
• Server functions you add you your Cloudflare Worker/Durable Object class.

A missing feature of Durable Objects is the ability to inspect and observe state outside of the JS API - you cannot access your data unless your write the HTTP endpoints yourself.

Durafetch:

• Implements those HTTP endpoints
• Implements a CLI to fetch the data from the endpoints, and then write them to a SQLite database

The two main use cases I have in mind are:

1. Development

   • When developing you need to refresh and run code paths repeatedly as you modify the code. Not being able to see what the state of the storage is makes this very difficult as you have to either print the values to a console, or write your own solution to read the state.
   • Even if you can read the state, you often need to filter/query it. SQL allows you to do this.

2. Production

   • Accessing and querying your server based state. E.g. How many sign-ups, sales. Usage based billing etc.

Why D1 cannot be used (for now)

Cloudflare have a SQLite based persistence API called D1 that looks good.

• D1 is in Alpha, and looks like it will be in GA in around 12 months which is too faraway.
• D1 is still distributed.
  • Each Durable Object gets its own SQLite instance, which means you will still need to join them into one central DB to do a global query.

Take a look

Durafetch is described in more detail at durafetch.madebyenzo.com or the Github repo.

Income Percentiles

• See https://income-percentiles.madebyenzo.com for details.
• SVG chart drawn directly without a framework to support custom features.

Bezier Curves

• See https://bezier-curves.madebyenzo.com for details.

People By Initials

The motive for this project is that there is currently no way to get a list of notable people via two character initials, which is essential to the DOMINIC system for memorizing two digit numbers in the book How to develop a perfect memory.

The people are listed by popularity. The data has been processed from Wikipedia using Python.

I also wanted to experiment with the Golang language from Google.

Lets Meet For Coffee

This was a side project I completed sometime in 2013 whilst living in Ho Chi Minh. The aim was to create a simple way for a prospect to arrange a quick meeting with me at a cafe that is convenient for them.

It is built on Google App Engine, Python, Backbone.js, Google Maps API and Bootstrap.