Project Notes

A virtual intercom service - ring a doorbell in your browser and connect directly via audio or video. No app install required.

A feature suggestion and voting system using instant-runoff based ranking to surface the most broadly preferred options across a group.

CompactList implements the List<Long> interface, but internally it uses a tree of variable word-width segments to improve performance and memory usage compared to an ArrayList. Performance tends to be worse for appends than an ArrayList but better for inserts. Memory usage is significantly reduced, even for incompressible random data where it approaches the memory use of an array of primitive longs (which happens to be the internal representation in this case). There are some performance metrics over at the GitHub repository. My aim is to use this for the internal index representation in CSView, which already uses an earlier version of this data structure. Source code at GitHub

A fast CSV viewer built for files that break everything else. Open multi-gigabyte files in seconds with background indexing, smooth scrolling through 300M+ rows, and modest memory usage - without ever modifying your data. CSView 2 is a ground-up native macOS rewrite optimised for Apple Silicon. CSView 1.x remains available for macOS, Windows, and Linux.

After reading this article about Magnetic Poetry I was taken by the idea that the fridge was no longer the central notice board of the home: Our age-old impulses haven't gone away—changes in technology have just encouraged us to take them elsewhere. "It used to be the fridge door was the center of the home," Kapell says. "That’s where people would hang their kids’ art, leave notes for each other, put pictures, all that stuff." So, I thought I would create a digital magnetic notice board, replete with fridge poetry, postcards and photos. I've got a few loose ends to tie up, but I'm pretty happy with the proof of concept.

A web tool for copying Asana projects between workspaces. Takes a complete snapshot of tasks, subtasks, attachments, custom fields, and assignees, then recreates them faithfully in any destination workspace. Free to use with no installation required. In maintenance mode as of April 2025. Full details Background The original tool started in 2013, based on a script posted to GitHub by AWeg and an earlier version by MhdHejazi. A web UI was added to help choose projects, then gradually updated to support new Asana features like Teams and Custom Fields. It became surprisingly popular - in no small part because Asana support would refer users looking to transfer projects between workspaces to it directly. Over six years it was sustained entirely by donations, just about covering running costs.

A tool for collating online grocery receipts and reporting on price and purchase trends over time. Lets you track what you spend, compare prices across shops, and spot patterns in your shopping history.

Task Queue Manager (TQM) is an AWS-native task queue service built on DynamoDB and Lambda, deployable via Terraform. Influenced by SQS, Celery, BullMQ, and other task queuing systems, it's designed for workloads where SQS falls short: specifically where you need fine-grained control over dispatch rate to downstream services, approximate priority ordering, or DAG-style task dependencies. SQS's Lambda integration handles polling for you, but this makes it hard to implement backpressure and circuit breaking without tuning concurrency and batch size - which is highly sensitive to Lambda execution time. TQM keeps you in control of dispatch, with rate limits defined by token buckets which can be used independently, or combined and shared by multiple queues. Design

Kevalin is a proof of concept distributed web platform built with modern web technologies. It allows contributors to publish content to the network without the need to host a website themselves, while maintaining control over their own content in the sense that the files making up their posts are freely downloadable from the network at any time, or can be marked for removal by participating peers. The network is not tied to any particular server, the content itself is identified by its hash and may be fetched from any participating peer. The core principle is that it must not require the installation of any additional software to join the network, nor require a user account for read-only access to public content, just like the traditional World Wide Web. There are several intended interaction modes for publishing and consuming different types of content. Tech stack

A content-addressed, deduplicated backup tool targeting cheap cloud blob storage. The design follows the approach pioneered by Brad Fitzpatrick's Brackup: large files are split into chunks using content-aware boundaries, each chunk is stored exactly once in a remote store indexed by its hash, and each backup is recorded as a manifest referencing those chunks. This makes every backup self-contained and independent while deduplication handles the incremental savings automatically. The practical goal is to make encrypted incremental backups to AWS S3 Glacier and similar services viable, with the added option of pooling deduplicated chunks across multiple machines or users in a single store, rather than paying for a separate backup for each.

An experimental multi-tenant CRM that bridges email and Asana for managing external customer contact. Asana's native email integration is unidirectional and designed for internal teams; Handl turns an Asana project into a shared inbox by accepting inbound emails via AWS SES, creating a task per thread, and routing follow-up messages as comments on the same task. The outbound half - replying to contacts as tasks are updated or commented on - is still in progress. The stack is serverless throughout: AWS SES for email, Lambda for processing, and DynamoDB for state and per-tenant config.

A distributed application framework for building streaming, event-driven pipelines on commodity serverless infrastructure - Lambda, ECS, Kubernetes, or anything else - while remaining language and platform agnostic. The core idea, drawn from experience with Apache Storm and Flink, is to separate the application logic (tasks invoked over a simple HTTP protocol) from the orchestration concerns (event queueing, dispatch, rate limiting, backpressure), handled by a sidecar layer that runs alongside each execution unit. The closest analogue is DAPR, a Microsoft Research project that launched around the same time and addresses a similar problem. Sprocket's emphasis was on the protocol layer: any service that can handle an HTTP request can participate in the graph, regardless of language or runtime.

Whalefish was an experiment to build a flexible cloud-native CMS based on a user-defined document taxonomy. Each document type has a number of fields, which may be content, metadata or references to other documents in the taxonomy. There's no hard-coded concept of a Post, the document structure could represent anything from a blog, to a product catalog, to a photo gallery. An initial version was started in 2016, built around similar technology to my blog and several other projects at the time: a Java web service implemented on Google App Engine, using the corresponding Google platform stack including Cloud Datastore. This attempt got fairly far, including document versioning, handling for multiple domains, preliminary hosting payment integration with Stripe, and an approval/publishing workflow. In the end it was the front-end elements which made the idea unworkable.

An experimental graph database written in Go, motivated by the poor performance of existing options like Neo4J and the question of whether a bare-metal approach could do better. The design used Morton-coded indexes to maintain bidirectional edge relationships between nodes in a way that could be distributed across a cluster, although it didn't progress that far. It reached a working single-node implementation with acceptable traversal performance, including some connected component algorithms, but never gained replication or clustering. In the end, confidence in the implementation wasn't high enough to use it in production, and the need for deep graph analysis turned out to be less pressing than expected. At Stashmetrics, the follower graph problem was solved more pragmatically by storing edge lists directly in DynamoDB - paginated arrays of follower IDs per user rather than one record per edge.

An implementation of 'Planning Poker' often used in software feature estimation. This version uses WebRTC data channels to coordinate votes between the players, which reduces the need for a server to track this state and improves running costs and performance.

A Twitter analysis tool that examined the lists a user appeared in and used that list-membership fingerprint to recommend accounts with a similar profile.

Languamatch was an experimental service for locating language exchange partners based on shared interests and skill levels. (Currently offline)

Furor is a web content curation tool for businesses. (Currently offline)

SlimTweet tries to shorten a message by applying a set of rules. For example, and will be replaced by &, while it is will be replaced by it's. Substitutions will be made one at a time until the message reaches a target length (140 characters by default). If SlimTweet still can't make the message fit, it will start tweaking the Unicode characters which make up your message, without changing its appearance too much. For example, the digraph vi will be replaced by the roman numeral character ⅵ. When rendered in a standard font, these substitutions are nearly inⅵsible. (Currently offline)

Stashmetrics was a seed-funded SaaS startup focused on social media analytics for content marketing, running from 2012 to 2017. The core problem it addressed was the difficulty of discovering what content was actually resonating within a specific niche community - as opposed to what was broadly popular - and finding influential accounts within that community. The two main capabilities were: 1. Niche content discovery - given a community of interest, surface the content and topics gaining traction specifically within that group, rather than across the platform as a whole. 2. Influencer similarity search - given a small seed set of influential profiles, locate similar accounts using profile text similarity and common list membership, combining graph analysis with vector similarity search. The platform targeted Twitter, Instagram, Facebook, YouTube, and Tumblr.

Communiscript, later renamed Retype, was a collaborative structured document editing system. The idea was to treat a technical document like version-controlled source code: break it into sections and paragraphs, version each of those individually, and let several authors work on the same document at once with comments and revisions attached to the specific paragraph they refer to rather than to the document as a whole. The original inspiration came from helping to write several academic papers, with each contributing author providing various sections and commenting on each other's parts. I was also providing a final pass for English language usage (in a multinational research group the range of native languages was quite wide!). This was accomplished by emailing drafts back and forth, and reconciling edits in MS Word, an unpleasant task to say the least. Communiscript, pre-2012

Graphviz is a widely used graph layout package developed by AT&T. There exist several ways of using Graphviz from Java, but in most cases the Graphviz command line tools are called to parse files containing a graph definition and render a rasterised image of the graph. This is unsatisfactory for many interactive applications, and a more direct interface to the layout algorithms is desirable. Java bindings generated using SWIG already exist as part of the Graphviz source code, however they are very low-level. jGraphViz aims to build on these low-level bindings to provide a more flexible interface to the graph model and layout system. It also provides some basic support for rendering the resulting layouts using GEF , and packages the library as an OSGi plugin compatible with Eclipse and other OSGi containers. Source code at SourceForge

In order to allow the use of Ada libraries with Java GUIs, a JNI (Java Native Interface) binding must be created which allows the JVM to call the non-java code, and vice versa. Working with the JNI is complicated, and maintenance to the interface is time consuming. A more intuitive approach is to make Java objects directly visible to the native code, and vice versa. Javada allows this with Ada by providing automatically generated 'proxy' objects, which present the correct interface to the developer while hiding the complexity of the JNI and performing memory management behind the scenes.

Wildebeest was a peer-to-peer distributed web cache, built in Java on top of ZeroC ICE middleware. The problem it aimed to solve was the Slashdot effect: before cloud hosting, a self-hosted project featured on Slashdot's front page would routinely go down or exhaust its bandwidth within hours. The idea was that participating users could share their locally-cached copy of a site through a distributed proxy, so availability would scale with popularity rather than being bottlenecked by a single host. The project reached proof-of-concept stage - a working proxy server that could serve cached resources from other nodes in the network, with the benchmark being successfully loading the Gmail web interface through it. It never had users and wasn't promoted.

A hobby x86 32-bit OS kernel built collaboratively with a housemate around 2005, inspired by a university OS design course. Written in D - chosen for its object-oriented model and optional GC, which made it a better fit than C++ for bare-metal development. The bootloader used ASM segments from the osdev.org wik; everything else was built from scratch, including a custom memory allocator and the D stdlib primitives needed in kernel mode (no libc available). The binary image was cross-compiled and loaded into QEMU. Design The kernel was designed as a microkernel built around non-blocking message passing - avoiding kernel interrupts for most IPC. Rather than a file-based abstraction, the design was object-based: OS primitives would expose interfaces that could be interacted with transparently over the network. This was influenced heavily by CORBA, which I had been working with during a placement year.

The Eclipse CORBA Plugin (ECP) provides extra features to the Eclipse platform to facilitate development using the CORBA middleware specification. Features IDL (Interface Definition Language) file editor with syntax colouring. Customisable syntax colouring Eclipse automatic installation site Tool for running external IDL compilers New Outline view for the IDL editor Autocompletion for editing IDL types Source code