04 / WORK
Listen. Understand.
Build.
Listening and understanding are part of what I do best. I bring them to my projects, my writing and my collaboration with AI. That’s how I give shape to the systems and the bonds that have a place in HME84.
Giving shape to ideas.
SYSTEMS / WHAT I BUILT
These systems grew out of problems I know first-hand: organising information, working with my photos and keeping projects moving. I provide direction and judgement; agents help turn that into code.
I define the problem, scope and acceptance criteria. Today, I delegate writing the code to agents and review what they deliver: implementation, tests and actual behaviour. If a delivery doesn’t solve the problem I set out, it goes back for revision. For me, orchestration means staying responsible for the work from start to finish.
KNOWLEDGE AND VOICE
Xerebro84
I needed to organise accumulated information and trace it back to its sources. The work began with the knowledge index at HME84 Tech Labs, followed by Cerebro, Cerebro Vivo and Knowledge Ledger. I decided to bring that work together in Xerebro84: a macOS app and a command-line tool. The requirement I set for the agents is to preserve provenance and make each answer verifiable. Its local search works without a network connection or an AI model; indexing documents doesn’t mean every question is answered.
2026-05 → 2026-10 · 11,205 documents indexed
Mox Talks
I wanted to talk to an assistant that could consult my knowledge, without starting from scratch in every conversation. Mox Talks connects voice to Xerebro84: listening and speech generation happen on-device, while the model can run locally or in the cloud. Working with the agents, context measurements revealed that some of the information sent to the model was being dropped; the window was adjusted on that evidence. Switching engines supports continuity, but it doesn’t make them equally capable.
2026-07 → 2026-10 · context window set to 8,192 tokens after measurements showed that 2,050 of 4,310 tokens were being dropped
MIRANDA
Miranda began in December 2025 as Apple Shortcuts to organise my photos: a technical bouncer scored focus, exposure, noise and intent; an editorial director classified and rated each image, then added a short phrase. Two partially working versions followed. In June 2026 I decided to start over with what I’d learned and separate personal use from the commercial branch. That decision guides the implementation work I delegate to agents and my review of what each part can actually do. Three paths emerged.
Moira
I made Moira the personal branch for reviewing my photos and making editorial decisions in a desktop app. Agents develop it; I review both how it behaves in use and its backend and interface tests. I use it personally: the documented tests help check changes, but they aren’t a promise that everything is finished.
2026-06 → 2026-10 · 286 backend tests and 452 interface tests
Miranda
I separated Miranda into a commercial branch for iPhone and iPad. My requirement for the agents is that its reading of a photo should rest on visible signals, with the evidence kept on-device. I check that intention against what the app delivers and its core tests. This branch is still in development; those tests alone don’t make it a product ready for public release.
2026-06 → 2026-10 · 210 core tests
Boîte Rouge
I wanted to bring the useful parts of those shortcuts into Photos. I asked the agents to rebuild the analysis, classification and album-creation workflow with one clear constraint: don’t delete or move the originals. Boîte Rouge applies that decision to sets of photos, and its catalogue records the images analysed. That volume shows how much has been processed, not how accurate every classification is.
2026-09 → 2026-10 · 11,055 photos analysed
WHERE I BRING TOGETHER WHAT I USE
ITL and WTL
I needed one place for the tools I use to manage work: finances, weekly operations, tasks and natural-language commands. ITL, HME84 Tech Labs, brings them together on the web and is in production. With the agents, I’m bringing that work into WTL, a desktop app with a local database. My review distinguishes what I already use in ITL from what has been validated in WTL: tests for one version don’t prove that both have all the same features.
2026-05 → 2026-10 · 283 tests in WTL
CONTINUITY AND JUDGEMENT
Sovereignty.
Before June 2026, we were already building a way of working that kept context, decisions and deliverables outside any single conversation. When I lost access to one of the models I used, that foundation let us continue: HME84’s infrastructure and way of working experienced no major disruption. A different model took over the role; I kept directing the work.
That episode made the orchestration we were already practising more explicit: assigning responsibilities, handing context to another agent and asking for evidence before accepting a result. In July I formalised it as a design principle: keep control of the knowledge, rules and work products, and be able to change providers while maintaining continuity.
That is what sovereignty means to me. I still use technology from different providers, with their limits on access, cost and capability. Changing models means checking what the replacement can do; continuity doesn’t guarantee identical answers. What I want to preserve is the ability to choose and keep working with what we’ve learned.
The dates trace each project’s development. The figures record documents, photos or tests; they don’t measure commercial impact or guarantee overall quality.