The Lab Notebook
A dated archive of what we built — often one to two years early.
This is the record: real R&D, dated, tiered by honesty, and honest about the failures — because a lab that hides its negatives forfeits the right to be believed on its results. Filter it, or read it as a timeline. Some reports are open; some are protected.
- 2024→2026dated, on the record
- 💡 101sparks (lightbulbs)
- ✅ 42proofs — each a hard figure
- 🔓 / 🔒public + protected
- 0faked results · nothing merely asserted
74 of 74 projects
💡 101 sparks · ✅ 42 proofs · 74 projects
33 Proven · 30 Partly proven · 6 Disproven · 0 Inconclusive · 5 Not run
The Zoe seat — a standing auditor whose job is to catch our own headline inflation
ProvenThe pitch said hallucination was impossible. The seat found the measured bound that pitch had quietly dropped, and put it back.
The recorded meter — why the mind tree retracted its own night scores
Proven5.5, 6.3, 6.4 looked like progress. They got retracted the same day, because none of them were ever saved as a run record.
The sibling letters — one tree mailed a proof, the other rebuilt it from source
ProvenThe proof of a proof isn't a citation. It's a second machine, given only the source, reproducing every table to the decimal.
The two-organ mind — deriving an architecture from a dead speed thesis
Partly provenThe micro model is the gate pointed at the mind — not a smaller version of it.
The navigator post-mortem — the honest death of a speed thesis
DisprovenTwo theses were supposed to make a 30B model fast. Measurement killed both — and that is the report worth publishing, not the one where they quietly survive.
PatentFactory — a working invention-council process, run across ten industries, that has filed zero patents
Partly provenA patentable-idea factory that actually ran: 13 council sessions, 75 ideas, 6 drafted provisional patents across 10 industries — and zero patents filed with any patent office. We correct our own earlier record to say so.
Safety profiles at compile time — the July state of a memory leak being a compile error
Partly provenA memory leak used to be a bug you found in testing. Now it's a line the compiler refuses to emit.
NovaOS — the synthesis of twelve donor trees, and the empty model slot we won't hide
Partly provenEighty requirements passed a tri-gate final review, a live monorepo persists real data, one package minted on a real blockchain — and the core model slot is still empty. That fact stays on the page.
QUANTA — from a GPU-native-language codename to a language that proves itself
Partly provenA language an AI writes, and a machine proves — the trust isn't in the code, it's in the agreement between two independent compilers.
The honesty ledger — a portable schema for not lying to yourself about progress
ProvenA patent only gets filed if a grep on the real codebase actually finds the claimed code. That one rule is the whole ledger, in miniature.
Hush — a wearable sensory organ that streams ternary bands, not video, proven on a two-node bench
Partly provenTwo ESP32 boards, one privacy-ring FSM, a heart-rate channel, and a camera feed encoded as ternary bands — streamed live to a laptop over Wi-Fi. The titanium shell and the $129 price are still a poster; the wiring underneath it already runs.
TRIAD — a ternary image/video codec that beats JPEG for real, and caught its own fabricated numbers on the record
Partly provenA codec built on {−,0,+} coefficients instead of DCT blocks beats JPEG by up to 43% on real photos — and the project's own results log contains a signed confession of three fabricated numbers, replaced with the real ones, on the public record.
NovaP/1 — a 16-byte header every Novaterra sensor organ agreed to speak, and a receiver that doesn't exist yet
Partly provenOne 16-byte header, one CRC, one magic number — NVP1 — that Hush's camera node streams live today. The spec calls itself 'the binding contract' precisely because the Rust receiver crate it's meant to anchor hasn't been written yet.
Council and gate — why unanimous 10/10 beats an average, and a pass has to exist on disk
ProvenUnanimous 10/10 or it's a failure — not an average, not a majority. That single rule is what stopped a 32-GPU fantasy from becoming a plan.
The Great Cull — subtracting 57 projects to close one membrane
ProvenEveryone else was adding. We deleted fifty-seven of our own projects in one day — reversibly, with a manifest — because you cannot converge what you refuse to prune.
The calculated mind — where the 'reason or refuse' thesis began
Not runCompute meaning instead of guessing it, and a system can only do two things with a question: answer it, or refuse. It cannot bluff in between.
Awesomekernel — one persistent GPU kernel, treated as a phase-locked oscillator
Not runOne kernel, never leaving the GPU for an entire training step — reframed as a phase-locked oscillator, with a counter that catches it if a fused stage silently does nothing.
Dolphin — 50 billion tokens, sharded, manifested, and handed over honestly
ProvenA tokenisation run died at shard 368 of 1,500 — and the handover letter says so, right next to the parts that are 100% verified.
FBL — reverse-compiling a real microservice into a spec, then regenerating it byte-identical
ProvenReal code, in, out as a spec, back out as code — byte-identical, across nine frameworks. That round-trip is the proof a code-generation demo can't fake.
GLYPH — compressing meaning, not text
Partly provenWhat if you compressed the meaning instead of the text? A dense language of 1,024 primitives that round-trips to English and back — and trains in under two GPU-weeks.
Newtraining — an honest plateau, then a real break, hunting for something not a transformer
DisprovenA hard-kill bar was named before the run started. The result didn't clear it — and that's exactly the kind of result this lab is built to publish.
Unification — the orchestration spine, and the incident it survived on the record
ProvenA 1-billion-parameter training run crashed out of memory mid-project — and instead of scrubbing the incident from the record, all three response rounds were preserved next to the final fix.
A model made of waves, not attention — the non-transformer architecture
Partly provenEveryone builds on attention. We built a sequence model out of wave equations — and made it prove it was actually running before we believed a single loss number.
A brain–computer interface for under £600 — no surgery, off published science
Partly provenNeuralink-class ambition, one-tenth the cost, zero surgery — a bidirectional brain–computer interface you could build for the price of a phone.
Magnus — the first attempt to unify a quarter-million lines, and the correction it forced
Partly provenBefore there was a synthesis, there was a first attempt — and the correction it forced (one real GPU, not a fantasy of thirty-two) propagated to every project that followed it.
A neural network on a 1982 ZX Spectrum — 2.2 KB, no floating point, no GPU
ProvenA real neural net, generating text, in 2,229 bytes on a computer older than most of the people reading this — no GPU, no floating point, no cloud.
NovaTerra Genesis V3 — six primitives for all of software, frozen at 70% built, 0% shipped
Partly provenAn internal review council rated the architecture 9.5 out of 10. The same record says 0% shipped. Both numbers stayed on the page.
PRISM — a GPU operating system, and datacentre-class inference on a single 3090
Partly provenWe built an operating system for the GPU — VRAM as a page table, models as processes — and pushed a real model to 147 tokens a second on a gaming card.
Trading Organism — where the council-review method was actually born
Partly provenA 28-module trading-research platform, still being wired together — but the real result is the review discipline it gave birth to.
A solar-farm SaaS on the most modern stack in the portfolio — council-reviewed UX
Partly provenGeospatial mapping, financial modelling, and 3D rendering, in one product, on the newest majors of every framework it touches.
NovaChain — a Cosmos-SDK fork to give AI agents a settlement layer, proven on devnet
ProvenAn agent economy needs the same tamper-evident settlement guarantees a financial system needs — so we built one, on a real (if small) running chain, not a spreadsheet.
NovaP — 15 bytes a token, and why the plumbing mattered more than the demo
ProvenNo demo, no screenshot — just 15 bytes a token instead of 200, and every token-throughput number upstream of it depends on that number holding.
BrainLang — a language for thought, not text, compiled to two lower targets
Partly proven768 primitives instead of a token vocabulary — roughly a 10.8x compression on the same content, with a lexer and parser that actually run.
Unification / Genesis — mapping 1,000+ SaaS categories onto one AI-native architecture
Not runAfter dozens of separate built projects, the moment we stopped building and mapped how one platform could, in principle, replace entire categories of software.
Strategic planning for a rail & metro operations product — documentation, not a shipped system
Not runTaking a technical product vision to market takes as much deliberate strategic work as the engineering build itself — this is that work, named honestly.
Rail PA/PIDS Platform — an 11-microservice control centre for metro PA and PIDS, built AI-assisted
ProvenEleven microservices, two hundred-plus endpoints, MQTT under the hood — proof that AI-assisted development can produce safety-critical software at real enterprise scale, not just prototypes.
The convergence platform — an orchestration tool built to help build itself
Partly provenA platform that generates software from natural language, judged first by whether it can help build itself.
Rail PIDS Project Monitoring — PostGIS for a region's scattered rail projects
ProvenDozens of scattered rail passenger-information projects across a region, one map — because a spreadsheet was never going to hold where they actually are.
GISGen — the gap between generating code and generating a working system
Partly provenGenerating code was never the hard part. Sequencing, wiring, and verifying it together as a working system was — and that's the gap GISGen went after.
A flying-club management system — specified in full, not (yet) built
Not runA specification thorough enough to be real work in its own right, in a domain — aviation training and licensing — that had never been touched before.
A network-LED monitor in C — the lowest-level production code in the whole portfolio
ProvenNot everything in this portfolio is a web app — this is direct, low-level hardware control in C, deployed and still running.
AIDE — running an LLM in the browser via WebGPU, and hitting the ceiling honestly
Partly provenWe put an LLM inside the browser tab itself, via WebGPU. It worked for small models — and told us exactly where the ceiling was.
The most enterprise-shaped stack in the portfolio — NestJS, queues, and a mobile client at once
Partly provenNestJS, a message queue, and a mobile client, all at once — the deliberate test of whether AI-assisted development scales past a single web app.
Three scripts, one UI — internationalisation from day one on a parking-guidance system
ProvenLatin, Tamil, and Chinese scripts, live in one UI, from day one — internationalisation done as architecture, not an afterthought.
The production agent hierarchy — Director → Manager → Worker → Tool
ProvenOne founder, a fleet of agents, and a hierarchy that plans without hallucinating — the orchestration engine behind the commercial work.
Version 24 — stewardship, not a rewrite, for a long-running car-park system
ProvenTwenty-four versions of a working .NET system, kept alive on purpose — proof that competent stewardship is sometimes the harder, better call than a rewrite.
A food & beverage consultancy site, delivered — animation and a business-analysis tool, end to end
ProvenDesign, animation, and a working business tool, delivered on a realistic client timeline — the unglamorous, real end of the archive.
RailMonitor — agents that read the railway industry so a bid team doesn't have to
ProvenInstead of a human scanning industry news for tender opportunities, agents do it continuously — and this is where vector search entered the portfolio for good.
TenderMS at 38% — what an overscoped integration looks like, honestly measured
Disproven38% complete, frontend never wired to the backend, running on mock data — stated plainly, because an honest failure is worth more to this archive than a polished half-truth.
Designer-quality PIDS screens with AI-assisted design — for a global transit programme
ProvenA functional-but-plain passenger display became designer-quality with AI in the loop — on a screen where legibility under pressure is a safety constraint, not a preference.
A full Linux terminal in the browser — WebContainers, no server anywhere
ProvenA real Node.js runtime, a real filesystem, and a real shell — all running inside a browser tab, with no server on the other end.
A browser that was an agent platform, not a window — built before Arc and Dia turned that way
Partly provenWe built a browser on the premise that the browser itself should be an agent — a place agents live and act — right as the industry started heading the same way.
CraftWorkz — keeping a second product line deliberately separate from the first
Partly provenTwo related AI product platforms, built in parallel, kept deliberately separate — the boundary was the lesson, not the code.
An isometric city builder in React — finished, playable, and small on purpose
ProvenA real, playable city builder, in React — proof that a UI framework built for forms and dashboards can hold a real-time game loop too.
Everything is an agent — and the agents write their own tools
Partly provenAn agent that hits a task it can't do usually fails. Ours wrote the skill it was missing and carried on — and every part of the platform, down to the gateway, was itself an agent.
TraderBot: Claude reasons about the market instead of following one fixed rule
ProvenThe same trading-bot idea from 2021, rebuilt with an LLM reasoning about several indicators at once instead of one fixed rule — the founder calls the difference night and day.
A private AI that lived on your own machine — before 'local AI' was a category
ProvenWe fine-tuned a model on one person's own browsing, entirely on their own machine, and asked it questions — a private AI, a year before that became the pitch everyone gives.
Incipio to CraftKit: 13+ iterations, and the real cost of finding the right module boundaries
ProvenThirteen-plus rewrites of the same platform's core architecture, each rename marking a real refactor of boundaries, not a coat of paint.
ChromeBot: a browser extension that acts, not just autocompletes
ProvenThe AI didn't just answer questions in the browser — it clicked, filled, and navigated inside one, on real pages, under real permission constraints.
Vibr: an AI site generator that worked — and the call to walk away from it anyway
Partly provenThe product worked. The market didn't wait. Knowing when not to keep building something turned out to be as much a skill as building it.
PicoAGI and PicoSaaS — the minimalism ideal, before the sprawl it didn't prevent
Partly provenTwo scaffolds, stripped to the minimum viable form, in the same week — the opposite instinct to the sprawl that, a year later, needed a cull to undo.
Grokit — a full Linux VM inside a Chrome extension, running AI-generated projects instantly
ProvenGenerate a web project with AI, then watch it run — inside a full Linux VM, inside a Chrome extension, with no server anywhere in the loop.
Daritana — a one-person, AI-assisted team out-featuring an enterprise PM tool
ProvenOne person, AI-assisted, built a PM platform that out-features a well-known enterprise incumbent — 120-plus features across three iterations, and we say plainly it's a scope claim, not an adoption claim.
An operating system, written by an AI — 50,000 lines of Rust that booted the wrong way
Partly provenWe asked whether an AI could write a whole operating system, not a snippet. It wrote fifty thousand lines of Rust and compiled clean — then got the boot mode wrong.
Hello Assembly: a 16-bit bootloader, written by hand, to see what sits beneath everything else
ProvenBefore directing an AI through a kernel's boot sequence, we did it ourselves, by hand, in 16-bit real mode — because you can't direct what you've never done.
Five versions in a month — the Chat Extension line and the habit it started
DisprovenFive versions of a browser chat tool in one month, in July 2023 — the direct ancestor of the '13 versions in 48 hours' pace AI tooling later made possible.
Dinosaur AGI — task decomposition, a year before 'agentic' was a word anyone used
Partly provenA year before 'agentic' was a buzzword, we built a system to break a task down and execute it — and learned that breaking it down is the hard part.
Animated Face: a speech-to-face pipeline, and a real lesson in real-time generative media
ProvenAudio in, a synchronised talking face out — and a hard early lesson about how unforgiving real-time inference is for generative media.
A control-room front end for a light-rail line — the gap between a demo and something operators trust
Partly provenThe first time React had to meet a control room, not a dashboard — where a wrong click has a real cost and glanceability isn't a nice-to-have.
GPT-3, a blog-post generator, and the moment code generation became the idea
Partly provenWe built a tool to draft blog posts. It quietly proved something bigger: if an LLM can draft a paragraph, it can draft a function.
First pass at rail tender automation — pricing and pack assembly, 2022
DisprovenThe first attempt at automating a genuinely painful manual process in the rail sector — tender pricing and pack assembly — and the first real lesson in how hard document generation actually is.
TweetDelete: a small, finished Django tool for bulk-deleting tweets
ProvenA small tool that did exactly one thing — delete tweets in bulk — and did it completely, end to end, with real OAuth and a real database behind it.
XMPP Messenger: a small, working rep of two processes talking in real time
ProvenA small XMPP client, built to answer one question: how do two processes actually talk to each other in real time?
CryptoTrader: EMA bots on Binance — the year fixed rules met a market that keeps moving
DisprovenFour strategy rewrites, one dashboard that survived all of them, and a lesson that took four years to fully land: the edge decays, the system doesn't.