Transparent by design

AI answers should show where they came from.

Language models are one layer of this portfolio's AI features — deliberately not the one in charge. Code decides what counts as evidence, retrieval finds it, a model writes about it, and validation checks the result against what was actually found.

Last reviewed: September 2026

The pipeline

How an answer is built

The Photography Curator and Engineering Ask follow the same order. Each stage is labelled with the kind of system that performs it.

Question
Performed by: Deterministic
Your question is length-limited, rate-limited and screened for attempts to redirect the assistant before anything else runs.
Retrieval
Performed by: Retrieval
Published records are found by exact terms and by meaning. A reranking model can reorder the strongest candidates.
Evidence
Performed by: Deterministic
A small, capped evidence set is assembled — for photographs, at most eight. A judgment model may admit a borderline photograph retrieval already found, never remove one. The set is the only portfolio material the writing model receives.
Writing
Performed by: Generation
A language model writes a structured answer from that evidence, citing what it uses.
Validation
Performed by: Validation
Citations must point into the evidence set, and structured claims are checked against stored data. A failure leads to one bounded retry, a removed citation or an explicit refusal.
Answer
Performed by: Deterministic
Every reply is labelled as coming from an AI assistant and shown beside the photographs or decisions it cites, so it can be checked against its sources.

Where the systems differ

Photo search stops after retrieval. Results are ordered by deterministic scoring; no language model writes anything.
Engineering Ask uses a model once before retrieval, to extract what a question is about — which is also how a German question is matched against the English records.
Follow-up questions are resolved by rules first. Only when a follow-up is ambiguous does a small model propose how it relates to earlier results — choosing among the same predefined plans, never touching evidence. If it fails, the rule-based reading is used.
Features that start from a photograph you are viewing, such as why two photographs are related, work from the selected photographs rather than typed text. Some involve no model at all.

Division of labour

Not everything is AI

Keeping decisions outside the language model is the point of the architecture: what can be computed exactly is computed exactly, and the model is asked to write.

Deterministic
Exact code with repeatable results.
Filters, ranking scores, relationships between photographs (same place, camera, colour or date), evidence limits, rate limits.
Retrieval
Stored records matched to your words, by exact term and by meaning.
Text and image embeddings, and a reranking model that reorders top candidates.
Generation
Written by a language model at the moment you ask.
Answer prose, the concepts behind an engineering question, readings of ambiguous follow-ups, spoken narration of a checked answer.
Validation
Checked by code before an answer is final.
Citations, structured claims, the answer's required shape, removal of leaked instructions.
Human-authored
Written or curated by me, stored, and not generated per visit.
Journal articles, engineering decision records, the portfolio's pages, the choice of published photographs.

Photograph descriptions and tags were drafted with AI assistance when each photograph was added. They are stored, editable metadata — not generated when you visit.

German wording of a cited engineering decision may be machine-translated for display. The English record remains the only evidence.

Systems

Different systems, different evidence

Each feature draws on its own sources, and not every feature uses a language model.

Photo search

DeterministicRetrieval

Evidence: Published photographs and their stored metadata.
Language model: None. Results are retrieved and ranked; nothing is written.

Photography Curator

DeterministicRetrievalGenerationValidation

Evidence: Up to eight published photographs with their metadata, camera details and the relationships computed between them.
Language model: Writes the answer from that evidence only. Its claims about the photographs are checked before they are shown.

Engineering Ask

GenerationRetrievalDeterministicValidation

Evidence: Engineering decision records I have written, plus material describing how I work.
Language model: Extracts the question's concepts, then writes the answer. Citations to decisions outside the evidence set are removed.

Engineering Journal

Human-authored

Evidence: Articles I have written.
Language model: None. The Journal is not generated, rewritten or summarised by AI on this site.

Boundaries

What AI does not do here

These limits are properties of the implementation, not of the prompt.

Reach beyond its evidence

The assistants cannot browse the web or read private, administrative or unpublished data. Each one receives only the evidence set built for your question.

Change the portfolio

The assistants have no write access to content. The portfolio changes only through authenticated administration.

Replace authored content

Journal articles and engineering decisions are written by me. Answers point to them; they don't stand in for them.

Become data

Answers are never stored as portfolio content or reused as evidence for later answers.

Act on its own

The assistants have no tools to call. The portfolio's MCP tools are limited to authenticated, owner-side document work.

Decide about people

The assistants answer questions about this portfolio. They do not assess visitors or make decisions about anyone.

When something goes wrong

No evidence: If nothing relevant is found, the assistant says so instead of answering from a model's general knowledge.
Failed checks: An answer that fails validation is retried once with the failure explained, reduced to what passes, or replaced by a fixed refusal.
Answers in progress: Text appears while it is being written. That provisional text is replaced by the checked answer, and narration only ever reads the checked version.
Provider outages: If the reranking service is unavailable, results keep their deterministic order. If an answer cannot be written, you see an error rather than an unchecked answer.
Variability: The same question can be worded differently from one visit to the next. The evidence and the checks stay the same; the phrasing does not.
Incomplete metadata: A detail visible in a photograph but missing from its stored description cannot be found or cited.

Implementation detail

Models, providers and your data

The underlying models may change. The evidence and validation contracts are designed not to depend on a particular model, so the current providers are listed here as implementation detail.

Anthropic (Claude)

Writes answers by default, extracts the concepts behind engineering questions and interprets ambiguous follow-ups.

Receives: Your question and the evidence set selected for it.

OpenAI

Turns text into embeddings so records and photographs can be matched by meaning. It never writes answers.

Receives: Your question text.

Cohere

Photography only: matches your words against the photographs' images by meaning.

Receives: Your question text.

OpenRouter

Routes four jobs to other models: reranking candidates; a yes-or-no judgment on whether a borderline photograph supports your question, which can only admit a photograph retrieval already found; speaking an answer when you ask for narration; and — where a feature is configured for it — alternative answer-writing models, with Claude as the final fallback.

Receives: Your question with short text descriptions of candidate photographs or records; for narration, the checked answer text.

What is kept

Spoken questions are transcribed by a speech model running in your browser. The recording is not uploaded; only the resulting text is sent, as your question.

Submitted questions are stored on this site so the assistants can be reviewed and improved. The record keeps the question, how it was answered and how long it took — and nothing that identifies you: no IP address, no cookie, no account, and nothing that links two questions to the same person. Stored questions are deleted automatically after 90 days.

This page makes no claim about how long the providers above retain what they receive, beyond what can be established from this repository and their applicable terms.

This page reflects the transparency principles of the European Union's AI regulatory framework. It provides transparency; it is not a legal certification or a claim of compliance.

European Commission: AI Act