AI Laboratory

Fourteen AI capabilities, including the four I have not built.

Organised by capability rather than by project, because that is the question an employer is actually asking. Each entry says what it is, how it works, and what it taught me — and carries a status that is enforced strictly.

Implemented & tested

8 — built, tested and in use. There is an artefact.

Partially built

2 — built in part, or built and not yet proven in the real world.

Planned — not built

3 — not built. No capability claim of any kind.

Why the unbuilt entries are here. A laboratory page listing only successes is a marketing page. RAG, document extraction and production deployment are common requirements in AI roles and I have not built them — so they appear with their status stated, linked to the roadmap step that will close each gap.

Capability area

LLM applications

Implemented & tested Hoshidub

Spoiler-safe episode companion

What it is

A conversational character that discusses a video the viewer is watching and is structurally unable to reveal what has not yet played.

How it works

The guarantee is architectural, not instructional. The model is only ever given the portion of the episode that has already been shown, so a request for the ending cannot be satisfied — there is nothing in context to leak. A prompt instruction would have been talked around on day one.

What it taught me

Where a behaviour genuinely must not happen, remove the capability rather than asking the model to decline. Policies are negotiable; absent context is not.

Implemented & tested Project Luna

Register control for a non-assistant voice

What it is

A live presenter that does not sound like a chatbot, despite being driven by an instruction-tuned model.

How it works

Personality expressed as a constraint on style rather than a long instruction, plus a filtering layer that rejects assistant phrasing before it reaches speech synthesis.

What it taught me

The default register of an instruction-tuned model is a persistent product problem. Post-generation filtering proved more reliable than asking the model to be someone else.

Capability area

AI agents

Partially built Project Luna

Unattended presenter with an autonomy loop

What it is

An agent that decides what to respond to, when to speak, when to stay silent and when to introduce a new topic, over a long session, with no operator input.

How it works

A director loop owns every decision and treats the language model as one input among several — chat events, silence duration, on-screen activity. Safety controls (emergency stop, moderation, per-viewer forget) were built before any thought of an audience.

What it taught me

Honest status: it runs unattended in testing and has never done so in public. The loop is built; the proof is not. That distinction is the whole reason this is marked partial.

Implemented & tested FRIEND//DEBT simulation engine

Enforced information boundary for reasoning agents

What it is

Eight simulated agents reasoning about each other under uncertainty, with an automated guarantee that none of them can see the hidden truth they are meant to deduce.

How it works

Knowledge is modelled with provenance — who knows what, from whom, when — and a test scans the agent-reasoning code for any access to ground-truth state and fails the suite if it finds one.

What it taught me

This is the same control as keeping a model away from its test set, and it belongs in a test rather than in a convention. A documented rule decays; a failing test does not.

Capability area

AI content evaluation

Implemented & tested Hoshi Nova

Automated narration verification

What it is

Synthesised narration is transcribed back by speech recognition and compared against the written script before an episode is allowed out of the pipeline.

How it works

Speech recognition used as an instrument rather than a feature. A synthesis dropout or a wrong take produces a transcript mismatch, which is detectable automatically; a human reviewing twenty hours of output will miss some, and will be bored long before the end.

What it taught me

Generated output needs a machine-checkable definition of correct. Without one, quality control is someone's attention span.

Implemented & tested

Pedagogical review: correct content that teaches the wrong thing

What it is

A review stage that asks what a reader would conclude, rather than only whether the facts are accurate.

How it works

It exists because a chart in instructional material I produced was entirely accurate and taught the opposite of its lesson: plotting the final value over a favourable sequence made the more aggressive configuration look better. Changing what the chart measured fixed it. No factual check would ever have flagged it.

What it taught me

The most dangerous failure in generated content is not inaccuracy, it is accurate material that implies something false. Catching it means evaluating the inference a reader will draw, not the facts on the page, which is what content evaluation work consists of.

Capability area

Model testing and evaluation

Implemented & tested

Split-half adoption testing

What it is

A rule that a change is adopted only if it independently improves both halves of a held-out split, with documented rejections.

How it works

Candidate changes are measured in-sample and out-of-sample separately. Several structurally appealing improvements were rejected for failing one half while improving the total — including one that produced more successful outcomes and performed worse overall.

What it taught me

Aggregate improvement hides overfitting. A held-out split with an adoption rule is the cheapest defence available, and the discipline is identical whatever is being evaluated.

Implemented & tested

Measurement-instrument validation

What it is

Establishing that the platform's own reported success metric was wrong for my system, and rebuilding the measurement before trusting any result.

How it works

A two-stage operation produces two platform records, inflating the reported count and success rate. Records are now paired back into whole outcomes before anything is measured. The change I eventually adopted moves the platform's reported figure down while the true figure goes up.

What it taught me

Validate the instrument before the hypothesis. Trusting the dashboard would have rejected a genuine improvement, and I would never have known.

Capability area

AI workflow automation

Implemented & tested Hoshi Nova

Unattended media production with a delivery gate

What it is

Twenty 4K episodes produced from structured data with no human in the loop during production, and an automated gate that can refuse to deliver.

How it works

Separately runnable stages, a resumable queue that outlives its controlling session, and a gate measuring resolution, frame rate and audio level that restores the previous master on failure.

What it taught me

Automating production without automating refusal just produces bad output faster. The gate is the part that makes the automation trustworthy.

Capability area

Document intelligence

Partially built Hoshidub

Subtitle-track parsing and schedule-driven synthesis

What it is

Parsing a structured timed-text document and using its timing to drive real-time speech synthesis ahead of each cue.

How it works

The subtitle track is the document; its cue times are the schedule. Lines are synthesised before they are needed, which is what produces zero delay.

What it taught me

Marked partial deliberately. This is structured-document parsing, not document intelligence in the sense an employer means it — no layout analysis, no extraction from unstructured or scanned sources, no OCR.

Planned — not built

Extraction from unstructured documents

What it is

Not built. Extracting structured fields from invoices, statements and reports with a confidence score and a human review path for low-confidence output.

How it works

Planned with a labelled evaluation set from the start, scored per field, because an extraction system without per-field accuracy figures cannot be trusted or improved.

What it taught me

Listed because it is a common requirement in AI operations roles and I want the gap visible rather than papered over.

Capability area

Retrieval-augmented generation

Planned — not built

RAG over my own project corpus

What it is

Not built. The intended first build: retrieval over my own technical documentation, decisions and rejected experiments, so the reasoning behind a past decision is answerable with citations.

How it works

Planned as chunking and embedding the corpus, a vector store, retrieval with citation of the source document, and an evaluation set scoring retrieval relevance and answer groundedness separately.

What it taught me

Named as a gap rather than implied as a capability. I have built context-constrained LLM features; I have never built a retrieval pipeline. It is step 6 on the roadmap and the evaluation harness matters to me more than the retrieval.

Capability area

Production AI deployment

Planned — not built

Containerised model serving with monitoring

What it is

Not built. Everything I have deployed runs either on a single workstation or on PHP shared hosting.

How it works

Roadmap steps 7 through 12: containerisation, a managed inference service, then monitoring for latency, cost, error rate and output drift.

What it taught me

The honest boundary of my current ability. I can build and verify an AI product on one machine; I have never operated one at scale, and no part of this site implies I have.

The pattern across all of it

Reading these back, the same idea keeps appearing in different clothes: where a behaviour genuinely must not happen, remove the capability rather than asking a model to decline.

The companion that cannot spoil an episode does not have the ending in its context. The reasoning agents that must not be omniscient are prevented by a test that fails the build. The pipeline that must not deliver bad output has a gate that restores the previous version instead. In each case the alternative was an instruction, a convention or a reminder — and all three of those are things I would eventually forget or a model would eventually talk its way around.

That is the most useful thing I have learned building AI products, and it generalises well beyond the specific systems here.

How I verify results The roadmap that closes the gaps