a11ybob.com

The model railway

Bob grew up in Northumberland, in the north of England, and built a series of model railways through his teenage years. One of them in particular is worth dwelling on. It was first controlled with relay logic and later rebuilt around TTL logic gates as parts became available to him through contacts that his school had with a local electronics company; a relay interlock is a real-time safety system in the formal sense, and the railway accordingly was a real-time control system in everything but name — direct, immediate, hardware-shaped.

Programming entered the same years from a different direction. In the late 1970s and early 1980s, Northumberland County Council ran an ICL 2900 mainframe and made batch-time on it available to secondary-school students; programs were written by hand on coding forms, sent in to be punched onto cards by typists, and run when the mainframe was quiet. Turnaround was measured in weeks, and a single typo could cost three of those weeks. Discipline, in that loop, was enforced by the loop itself. Two modes of working with logic in parallel, then — the real-time physical immediacy of the railway, and the slow deliberate abstraction of batch programming. Everything that came after is a variation on what was learned in those years.

By early 1981, Bob and his father had kit-built a ZX80 on the kitchen table at home. A 16K Spectrum followed, and on it Bob learned Sinclair Spectrum Forth — a stack-oriented language that ran close to the metal because the metal was all you had.

That model returned to his writing, in his doctoral work. The thesis chapter on haptic design names it openly: the inspiration for this model of haptics is relay logic. I was using relays in the 1970s for my O-level Engineering project to control a model railway. Same railway, same relay logic, repurposed as the foundation for a formal model of haptic interaction. Not a metaphor reached for in retrospect; a direct acknowledged inspiration, named in the thesis itself.

Teesside, and any hardware that fitted in three months

Bob read BSc Computer Technology at Teesside Polytechnic from 1982 to 1986 — a four-year course split evenly between hardware and software. During his time there the polytechnic decommissioned its Univac 1100 mainframe; the building it had occupied was converted into two squash courts. The conversion dates the era for him specifically, in a way that more abstract dating could not.

From 1984 he was working at Metal Box Automation and Controls in Worcester, in a culture summed up in one sentence: “Any hardware, any language, any architecture — so long as projects were kept to tight deadlines, typically three months.” Three projects from those five years are worth naming. The first was the programmable logic controllers that ran the Cadbury Cream Egg production line, written in ladder logic with relay-symbol diagrams. The second was the Metal Box Microguage, an industrial quality-control system for measuring the wall thickness and height of beer cans — factory-floor version in 8085 assembler, portable briefcase version in PolyForth on a PC. The third was the Difco automated blood culture system, initially prototyped on Transputers — chosen for their built-in networking and an elegant fit for the problem — running occam, with the production version moved to the less exciting networked 8051 microcontrollers in each drawer; first generation of what is now standard in pathology labs. The work was practitioner-level rather than algorithmic; what mattered was finding an elegant solution, not a novel one. Originally aimed at the food industry, where it turned out to be too slow and too expensive, the Difco system found its market in medicine, where it was neither.

The briefcase Microguage carried a teachable parable. The earlier version had been implemented in floating point, and its rounding errors had been making its readings disagree with the printed reports the factory had been working from for years. Bob rewrote it in fixed point — the natural Forth idiom — at a precision that matched what the underlying test equipment could actually deliver. The readings became repeatable. One customer, presented with both the corrected values and the more honest precision, accused him of cheating. The accusation was wrong but it was instructive: the gap between accuracy and repeatability — and between the precision a tool claims and the precision its instruments can support — is the gap between a measurement people trust and one they do not. The same gap shows up in accessibility testing twenty years later.

Distributed systems, switching, adaptive UIs

Four projects from this period, in chronological order:

An armoured tracked vehicle in desert paint with a missile-launcher turret carrying twin missile pods and a sensor pod, mounted on an APC-style chassis. — Low Level Air Defence System (LLADS) — one of the Canadian Forces communications-systems contracts at Siemens Plessey Defence Systems, Havant, from 1989.

Network diagram showing five Unix nodes — two Secret, one Top Secret, two Unclassified — each connected through a Trusted Network Interface Unit (TNIU) to a shared LAN. Labelled 'Figure 3: A securely partitioned system.' — The Trusted Network Interface Unit (TNIU) at RSRE, extending the OSI seven-layer model with security layers and specified in Z notation.

A dark desk telephone with a wide LCD display, rows of function and numeric keys, and a slide-out QWERTZ keyboard underneath, photographed next to its product manual. — The Ascotel Crystal — an Ascom PABX terminal of the period. Software Group Leader for the introduction of object-oriented technology at Ascom AG, Solothurn, 1993–1996.

Architecture diagram of Control Data Systems Mail*Hub: proprietary endpoints (cc:Mail/Novell, Lotus Notes, IBM OfficeVision) on the left feed a Mail*Hub routing core containing a translation engine, addressing engine and message queuing, which in turn feeds out to the global OSI network on the right — SMTP gateway (RFC 1327), X.400 Message Transfer System, and the X.25 packet-switched data network. — The Control Data Systems Mail*Hub architecture, 1996 — the “giant giant Perl program” bridging Microsoft Mail, Lotus, IBM systems and X.400 through an X.500 directory.

From 1989 Bob was at Siemens Plessey Defence Systems in Havant, working on communications systems for clients that included the Canadian Forces here in Canada. A stint at the Royal Signals Research Establishment in the same period extended the OSI seven-layer model with security layers and built the Trusted Network Interface Unit, specified in Z notation. I can read Z. I can never write efficient Z.

GEC Plessey Telecoms followed in 1993, then Ascom AG in Solothurn from 1993 to 1996, where Bob was Software Group Leader for the introduction of object-oriented technology to Ascom’s switching business — ISDN telephony and PABX systems, including specialist exchanges for radio call-in shows. The notation methods of the era ran through everything: Shlaer-Mellor, Executable UML, SDL.

Then in 1996 came the Control Data mail hub, which bridged the pre-SMTP fragmentation of corporate email — Microsoft Mail, Lotus, IBM systems, X.400 — by converting between formats and routing through an X.500 directory. A giant giant Perl program, a write-only language as we used to say, which was a bit scary. A year of dynamic crew planning at British Airways followed.

Between 1997 and 2001, Bob was at Nokia Mobile Phones in Finland and Germany, working on a joint project with Nokia Research on adaptive UI design methodologies for novel interfaces in in-car telephony. The specific goal was adapting a phone’s UI for a third-party in-car infotainment system — the same phone software, adapted for whichever platform was hosting it. This is where adaptive UIs entered the working vocabulary, several years before the accessibility application emerged.

The shift to accessibility

In the early 2000s, three family motivations layered. Bob became increasingly involved in his cousin’s welfare as the MS progressed, and that involvement became the technical focus of much of the doctoral work that followed — see TUP, the first of the tools built specifically for him. Bob’s mother has arthritis, and her experience of typing on mobile devices informed the input-reduction direction of the same tool. Then, late in the same period, Bob’s father’s final illness put the family in an ICU where picture-board AAC proved inadequate for a man trying to participate in his own treatment decisions. None of these alone is the formal cause of the research focus, but together they turned the technical work toward accessibility, and they kept it concrete — built for specific people, not for an abstract user category.

Bob enrolled at Teesside two days a week for an MSc in Multimedia Applications, with a thesis on mobile accessibility for people with MS; the MSc became a PhD candidacy. There is also a third named user the framework was built for and tested against: the author himself, who has mild deuteranomalia, disclosed in the 2009 User Capability in an Adaptive World paper. The researcher is inside the design space he is modelling.

The doctoral research ran from 2004 to 2013. It began on the effectiveness of self-adapting user interfaces as assistive technology in hand-held mobile devices, and over those years grew into formal models of UI adaptation and of user capability. Most of the writing was completed and several peer-reviewed papers published from the work, including the W4A 2008 paper The CISNA Model of Accessible Adaptive Hypermedia(with Steve Green and Elaine Pearson). The viva was never taken — largely because of a family move to Canada, and CNIB and CELA happening. Two pieces of original conceptual work survived from the period and are still load-bearing for current writing: the CISNA Model itself, and Tetris as accessibility testbed, which turned out to be the most teachable artefact of the entire period — and produced an audio rendering that re-framed the modality question itself. Beneath both, the Measure of Accessibility collection — six pages from the Defining Accessibilitychapter — gives the formal-and-political theory.

A period in Beijing around 2004 sits inside the doctoral years. Bob’s husband, Taodi, had finished his Masters and joined Nokia in Beijing; the route to a same-sex partner visa ran through Tsinghua University, where Bob spent two semesters studying Mandarin so he could live in China while Taodi worked. The gruelling 6am bus journey to Tsinghua University produced the User Interface as an Evolving Community of Practice chapter — one of the more biographically rooted pieces of the corpus, written largely on that bus. The Communities of Practice framing went on to become the substrate for the 2029 framework.

A period in Singapore in the mid-2000s, also inside the doctoral years, brought connections to the gay Deaf community there. Out of a challenge over drinks at the bar Tantric on a Friday night came Sign16— a sign-writing system that fit on the 16-key keypad of a feature phone. The price of solving the problem was a beer. That story is the first of three on the Spotlight projects page.

In 2008, Bob and Taodi moved from the UK to Canada when Taodi joined BlackBerry in Waterloo. The PhD candidacy continued formally for another five years; the centre of gravity shifted.

CNIB, CELA, and the working tools

From 2013, Bob’s first role at CNIB was maintaining the original CNIB Library website whilst planning and developing its replacement, CELA, the Centre for Equitable Library Access — the moment Canadian public libraries finally took on the responsibility for accessible book access that CNIB had carried for a century. From 2019, as Manager of Digital Accessibility at CNIB Frontier, Bob built out the audit processes (websites, apps, lived user testing, electronic documents), the toolchain that supported them, and an early automated testing tool; the role also brought him into contact with the W3C ACT-R community group defining standard tests for digital accessibility, influencing his own work on accessibility testing.

A spell from 2021 as Director of Delivery at Digita11y Accessible Inc. covered the audit department of a SaaS-era consultancy — an in-house team with external agency-supplied subcontractors, audit-reporting and test-automation tools, and training and consultancy for George Brown College, Centennial College, Co-operators, Brookfield Properties, and Philip Morris.

The current role, Head of Accessibility at CNIB Access Labs, is the strategic technical seat at CNIB’s commercial accessibility arm and the methodology-owner position for the practice as a whole. Three production tools are in active use at Access Labs: autoA11y, the commercial accessibility-testing platform; Dictaphone, an AI-assisted audio-and-video accessibility analyser working from lived-experience and audit recordings, in production for over two years; and pdfMax, a PDF accessibility platform that tests every document against 122 distinct accessibility standards spanning PDF/UA, WCAG 2.2, the Matterhorn Protocol, and PDF/UA-1. The Three Pillars the public deck names — automated testing, manual inspection, lived experience — integrate through those tools into a single audit pipeline. All three are CNIB-owned and are mentioned here as practice, not portfolio.

In parallel, and on personal time, Bob develops three open-source tools of his own. Paradiseis a multi-model accessibility analyser whose technical contribution is the ActionLanguage intermediate representation (IR) — a JavaScript-to-semantic-tree transcoder that lets analysis reason about runtime behaviour at the source level. Carnforthis a Chrome DevTools extension testing the accessible-name computation algorithm in depth, the runtime counterpart to Paradise’s source-level work. automated-testingis a set of five AI-driven proofs-of-concept targeting classes of issue commercial scanners cannot reach. A fourth open-source line is in development: vision-AI analysis of screen recordings of user interaction using Qwen 3.5 (Alibaba’s vision-language model, via API), as the video/vision counterpart to Claude-based audio analysis in the audio-and-video pipeline.

Running alongside the testing tools is a seven-year sustained body of work on accessible maps: three working demos across building, subdivision, and city-neighbourhood scales, sharing one SVG-tile architecture, with a paper-shaped theoretical contribution — spatial cognition under modality conversion— that names the polar-coordinate finding the maps work produced.

The pattern across these is one Bob describes directly: “I play at home, and some of that work makes its way through to CNIB. They’re my employer, and a charity I believe in, and I love gifting them apps and code.” Both autoA11y and Dictaphone began as personal projects with roots in the PhD, and were later gifted to CNIB — Carnforth GPL became autoA11y, the home audio-analysis PoC became Dictaphone. pdfMax originated within Access Labs itself, but on the legacy those two earlier gifts had established. The home/CNIB asymmetry isn’t policy; it’s preference. Hobby code is the right shape for theoretical exploration, and production code is the right shape for institutional users with conformance obligations. Both keep happening.

The ActionLanguage IR in Paradise today carries forward the Forth-style threaded-interpreter execution model from the doctoral Action Language, which carried it forward from PolyForth at Metal Box in 1984. Same execution model, four problem domains, four decades.

The 2029 project

In 2006, during the doctoral work, Bob built a theoretical framework that treats accessibility not as a binary pass/fail and not as a static property of either system or user, but as an emergent dynamic equilibrium in a contextual space of competition between environmental factors, technical constraints, user capability, user preference, and the available UI resources and modalities. The framework requires agency for each factor: each must be able to advocate, negotiate, and respond.

The work was not abandoned in 2013; it was paused. The thesis conclusion explicitly named the next research step: that competition would be best considered through the use of game theory, with the mathematical formalism likely to be found in models of autonomous agents. In 2006 the theory was understood but the compute was not there. In 2013 the formalism was clearer but autonomous-agent implementations remained academic toys. With practical agentic AI now available, the implementation tools have caught up at last. The 2029 project is the resumption of named research, not a hobby waiting for retirement. Read on.